sensemaking 0.7.2 → 0.9.0

package/dist/esm/features/embed.js.map

@@ -1 +1 @@
-{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/embed.ts"],"sourcesContent":["import { existsSync, mkdirSync, readFileSync, renameSync, writeFileSync } from 'node:fs';\nimport { homedir } from 'node:os';\nimport { join } from 'node:path';\nimport type { DatabaseSync } from 'node:sqlite';\nimport type { Config, ResolvedConfig } from '../config.ts';\nimport { embedConfig } from '../config.ts';\nimport { SenseError } from '../errors.ts';\nimport { parseFile } from '../scan.ts';\nimport type { Feature } from './types.ts';\n\n// embeddings(path, chunk, start_line, end_line, scale, vector): heading-based chunks,\n// int8 vectors with a per-vector dequantization scale, vector NULL = not yet embedded.\n// Reconcile stores dirty rows inside its transaction; embedding tops up on the next\n// semantic query (embedPending), so staleness costs recall, never correctness.\n\n// Storage lever fixed by the bake-off (BENCHMARKING.md): int8 at 256 dims is\n// quality-free vs f32-512 when fused. Queries stay f32 at the same dims.\nconst STORE_DIMS = 256;\nconst BATCH = 64;\n\nexport interface EmbedProvider {\n  id: string; // model identity; participates in the cache key, change -> re-embed\n  dims: number;\n  embed(texts: string[]): Promise<Float32Array[]>;\n}\n\ninterface Chunk {\n  startLine: number;\n  endLine: number;\n  text: string;\n}\n\n// Deterministic chunker used at reconcile (line ranges stored) and at embed time (text\n// re-derived from the file -- chunk text is never stored). Heading-delimited with the\n// preamble kept; whole file when no headings. The title/summary prefix mirrors the\n// bm25 column weighting.\nfunction chunksOf(raw: string, search?: { title: string; summary: string }): Chunk[] {\n  const lines = raw.split('\\n');\n  const starts: number[] = [];\n  let inFence = false;\n  for (let i = 0; i < lines.length; i++) {\n    if (/^(```|~~~)/.test(lines[i])) {\n      inFence = !inFence;\n      continue;\n    }\n    if (!inFence && /^#{1,6} +/.test(lines[i])) starts.push(i + 1);\n  }\n  const bounds = starts.length === 0 ? [1] : starts[0] > 1 ? [1, ...starts] : starts;\n  const prefix = [search?.title, search?.summary].filter(Boolean).join('\\n');\n  const chunks: Chunk[] = [];\n  bounds.forEach((start, i) => {\n    const end = i + 1 < bounds.length ? bounds[i + 1] - 1 : lines.length;\n    const text = lines\n      .slice(start - 1, end)\n      .join('\\n')\n      .trim();\n    if (text.length > 0) chunks.push({ startLine: start, endLine: end, text: prefix ? `${prefix}\\n${text}` : text });\n  });\n  return chunks;\n}\n\nexport const embed: Feature = {\n  name: 'embed',\n  schema(db) {\n    db.exec(`CREATE TABLE IF NOT EXISTS embeddings (\"path\" TEXT, chunk INTEGER, start_line INTEGER, end_line INTEGER, scale REAL, vector BLOB, PRIMARY KEY (\"path\", chunk))`);\n  },\n  extract(raw, _body, search) {\n    return chunksOf(raw, search);\n  },\n  remove(db, path) {\n    db.prepare('DELETE FROM embeddings WHERE \"path\" = ?').run(path);\n  },\n  store(db, path, extracted) {\n    const insert = db.prepare('INSERT INTO embeddings (\"path\", chunk, start_line, end_line, scale, vector) VALUES (?, ?, ?, ?, NULL, NULL)');\n    (extracted as Chunk[]).forEach((c, idx) => insert.run(path, idx, c.startLine, c.endLine));\n  },\n};\n\n// --- static type: Model2Vec safetensors + pure-JS tokenizer, model cached in ~/.cache ---\n// Encode convention from model2vec/model.py: no special tokens, drop unk ids, mean-pool,\n// L2-normalize.\n\nfunction fetchToFile(url: string, dest: string): Promise<void> {\n  return fetch(url).then(async (res) => {\n    if (!res.ok) throw new SenseError('EMBED_MODEL', `model download failed: ${url} -> HTTP ${res.status}`);\n    writeFileSync(`${dest}.part`, Buffer.from(await res.arrayBuffer()));\n    renameSync(`${dest}.part`, dest);\n  });\n}\n\nasync function staticProvider(model: string): Promise<EmbedProvider> {\n  let dir = model;\n  if (!existsSync(join(dir, 'model.safetensors'))) {\n    dir = join(homedir(), '.cache', 'sensemaking', 'models', model.replace(/\\//g, '--'));\n    mkdirSync(dir, { recursive: true });\n    for (const file of ['model.safetensors', 'tokenizer.json']) {\n      if (!existsSync(join(dir, file))) {\n        console.error(`fetching ${model}/${file} into ${dir} (once; delete to refetch)`);\n        await fetchToFile(`https://huggingface.co/${model}/resolve/main/${file}`, join(dir, file));\n      }\n    }\n  }\n\n  const raw = readFileSync(join(dir, 'model.safetensors'));\n  const headerLen = Number(raw.readBigUInt64LE(0));\n  const header = JSON.parse(raw.subarray(8, 8 + headerLen).toString('utf8')) as Record<string, { dtype: string; shape: number[]; data_offsets: number[] }>;\n  const entry = Object.entries(header).find(([k]) => k !== '__metadata__');\n  if (!entry || entry[1].dtype !== 'F32') throw new SenseError('EMBED_MODEL', `${model}: expected an F32 safetensors matrix`);\n  const spec = entry[1];\n  const dims = spec.shape[1];\n  const dataStart = raw.byteOffset + 8 + headerLen + spec.data_offsets[0];\n  const matrix = dataStart % 4 === 0 ? new Float32Array(raw.buffer, dataStart, spec.shape[0] * dims) : new Float32Array(raw.buffer.slice(dataStart, dataStart + spec.shape[0] * dims * 4));\n\n  const tokenizerJson = JSON.parse(readFileSync(join(dir, 'tokenizer.json'), 'utf8'));\n  // Lazy import: the tokenizer loads only on the semantic path, never at CLI startup.\n  const { Tokenizer } = await import('@huggingface/tokenizers');\n  const tok = new Tokenizer(tokenizerJson, {});\n  const unkId = tokenizerJson.model?.vocab?.[tokenizerJson.model?.unk_token] ?? -1;\n\n  function one(text: string): Float32Array {\n    // The tokenizer yields undefined (not the unk id) for tokens outside the vocab; an\n    // undefined id would index the matrix at NaN and poison the whole mean-pool -- and the\n    // int8 conversion then stores the NaN vector as all zeros, silently. Keep integers only.\n    const ids = (tok.encode(text, { add_special_tokens: false }).ids as number[]).filter((id) => Number.isInteger(id) && id !== unkId);\n    const v = new Float32Array(dims);\n    if (ids.length === 0) return v;\n    for (const id of ids) {\n      const off = id * dims;\n      for (let d = 0; d < dims; d++) v[d] += matrix[off + d];\n    }\n    let norm = 0;\n    for (let d = 0; d < dims; d++) {\n      v[d] /= ids.length;\n      norm += v[d] * v[d];\n    }\n    norm = Math.sqrt(norm) + 1e-32;\n    for (let d = 0; d < dims; d++) v[d] /= norm;\n    return v;\n  }\n\n  return { id: `static:${model}`, dims, embed: async (texts) => texts.map(one) };\n}\n\n// --- api type: one POST against any OpenAI-compatible /embeddings endpoint ---\n\nasync function apiProvider(model: string, url: string | undefined, keyEnv: string | undefined): Promise<EmbedProvider> {\n  if (!url) throw new SenseError('EMBED_MODEL', 'features.embed.type \"api\" requires a url');\n  const base = url.replace(/\\/+$/, '');\n  const headers: Record<string, string> = { 'content-type': 'application/json' };\n  const key = keyEnv ? process.env[keyEnv] : undefined;\n  if (key) headers.authorization = `Bearer ${key}`;\n\n  async function post(texts: string[]): Promise<Float32Array[]> {\n    const res = await fetch(`${base}/embeddings`, { method: 'POST', headers, body: JSON.stringify({ model, input: texts }) });\n    if (!res.ok) throw new SenseError('EMBED_MODEL', `${base}/embeddings -> HTTP ${res.status}`);\n    const body = (await res.json()) as { data: Array<{ embedding: number[] }> };\n    return body.data.map((d) => Float32Array.from(d.embedding));\n  }\n\n  const dims = (await post(['dimension probe']))[0].length;\n  return { id: `api:${base}:${model}`, dims, embed: post };\n}\n\nconst providers = new Map<string, Promise<EmbedProvider>>();\n\nfunction getProvider(cfg: Config): Promise<EmbedProvider> {\n  const e = embedConfig(cfg);\n  if (!e) throw new SenseError('EMBED_DISABLED', 'semantic expansion needs features.embed in sense.config.json (e.g. \"features\": { \"embed\": true })');\n  const sig = `${e.type}:${e.model}:${e.url ?? ''}`;\n  let p = providers.get(sig);\n  if (!p) {\n    p = e.type === 'api' ? apiProvider(e.model, e.url, e.key) : staticProvider(e.model);\n    providers.set(sig, p);\n  }\n  return p;\n}\n\n// Slice + re-normalize (Matryoshka); optionally round through int8 storage.\nfunction toStore(full: Float32Array, dims: number, int8: boolean): { v: Float32Array; scale: number } {\n  const v = new Float32Array(dims);\n  let norm = 0;\n  for (let d = 0; d < dims; d++) norm += full[d] * full[d];\n  norm = Math.sqrt(norm) + 1e-32;\n  for (let d = 0; d < dims; d++) v[d] = full[d] / norm;\n  if (!int8) return { v, scale: 1 };\n  let max = 0;\n  for (let d = 0; d < dims; d++) max = Math.max(max, Math.abs(v[d]));\n  return { v, scale: max / 127 || 1 };\n}\n\n// Embed rows whose vector is NULL, re-deriving chunk text from the files through the\n// same parse + chunker that stored the rows.\nexport async function embedPending(db: DatabaseSync, cfg: Config, baseDir: string): Promise<void> {\n  const provider = await getProvider(cfg); // throws EMBED_DISABLED before touching the table\n  const dirty = db.prepare('SELECT \"path\", chunk FROM embeddings WHERE vector IS NULL ORDER BY \"path\", chunk').all() as Array<{ path: string; chunk: number }>;\n  if (dirty.length === 0) return;\n  const storeDims = Math.min(STORE_DIMS, provider.dims);\n\n  const byPath = new Map<string, number[]>();\n  for (const row of dirty) {\n    const list = byPath.get(row.path) ?? [];\n    list.push(row.chunk);\n    byPath.set(row.path, list);\n  }\n\n  const jobs: Array<{ path: string; chunk: number; text: string }> = [];\n  for (const [path, chunkIdxs] of byPath) {\n    let chunks: Chunk[];\n    try {\n      chunks = parseFile({ relPath: path, absPath: join(baseDir, path), mtimeMs: 0, size: 0 }, [embed]).doc.extracted.embed as Chunk[];\n    } catch {\n      continue; // vanished since reconcile; the next reconcile removes its rows\n    }\n    for (const idx of chunkIdxs) if (chunks[idx]) jobs.push({ path, chunk: idx, text: chunks[idx].text });\n  }\n\n  const update = db.prepare('UPDATE embeddings SET scale = ?, vector = ? WHERE \"path\" = ? AND chunk = ?');\n  for (let i = 0; i < jobs.length; i += BATCH) {\n    const batch = jobs.slice(i, i + BATCH);\n    const vectors = await provider.embed(batch.map((j) => j.text));\n    db.exec('BEGIN');\n    try {\n      batch.forEach((job, j) => {\n        const { v, scale } = toStore(vectors[j], storeDims, true);\n        const q = new Int8Array(storeDims);\n        for (let d = 0; d < storeDims; d++) q[d] = Math.round(v[d] / scale);\n        update.run(scale, Buffer.from(q.buffer), job.path, job.chunk);\n      });\n      db.exec('COMMIT');\n    } catch (err) {\n      db.exec('ROLLBACK');\n      throw err;\n    }\n  }\n}\n\n// Top candidates by cosine for a semantic find: best chunk per file, its line range\n// riding along. FTS5 operators in the terms are lexical syntax, not meaning -- stripped\n// before embedding.\n// Returns the cosine similarity alongside each candidate: stored and query vectors are both\n// L2-normalised before quantisation, so the dot product is cosine in [-1, 1]. `find` surfaces\n// it because the fused rank score cannot express match quality -- nearest-neighbour search\n// always returns a nearest neighbour, so without a magnitude an agent cannot tell a real hit\n// from the best of a bad lot.\nexport async function semanticCandidates(db: DatabaseSync, cfg: Config, terms: string, fetch: number): Promise<Array<{ path: string; lines: string; similarity: number }>> {\n  const baseDir = (cfg as Partial<ResolvedConfig>).baseDir;\n  if (!baseDir) throw new SenseError('EMBED_MODEL', 'semantic expansion needs a config with baseDir (use loadConfig/open)');\n  await embedPending(db, cfg, baseDir);\n\n  const provider = await getProvider(cfg);\n  const storeDims = Math.min(STORE_DIMS, provider.dims);\n  const text = (terms.match(/[\\p{L}\\p{N}]+/gu) ?? []).filter((t) => !['AND', 'OR', 'NOT', 'NEAR'].includes(t)).join(' ');\n  const { v: qv } = toStore((await provider.embed([text]))[0], storeDims, false);\n\n  const rows = db.prepare('SELECT \"path\", start_line, end_line, scale, vector FROM embeddings WHERE vector IS NOT NULL').all() as Array<{\n    path: string;\n    start_line: number;\n    end_line: number;\n    scale: number;\n    vector: Uint8Array;\n  }>;\n\n  const best = new Map<string, { score: number; lines: string }>();\n  for (const row of rows) {\n    const q = new Int8Array(row.vector.buffer, row.vector.byteOffset, Math.min(storeDims, row.vector.byteLength));\n    let dot = 0;\n    for (let d = 0; d < q.length; d++) dot += q[d] * qv[d];\n    const score = dot * row.scale;\n    const existing = best.get(row.path);\n    if (!existing || score > existing.score) best.set(row.path, { score, lines: `L${row.start_line}-${row.end_line}` });\n  }\n  return [...best.entries()]\n    .sort((a, b) => b[1].score - a[1].score)\n    .slice(0, fetch)\n    .map(([path, b]) => ({ path, lines: b.lines, similarity: Math.round(b.score * 1000) / 1000 }));\n}\n"],"names":["existsSync","mkdirSync","readFileSync","renameSync","writeFileSync","homedir","join","embedConfig","SenseError","parseFile","STORE_DIMS","BATCH","chunksOf","raw","search","lines","split","starts","inFence","i","length","test","push","bounds","prefix","title","summary","filter","Boolean","chunks","forEach","start","end","text","slice","trim","startLine","endLine","embed","name","schema","db","exec","extract","_body","remove","path","prepare","run","store","extracted","insert","c","idx","fetchToFile","url","dest","fetch","then","res","ok","status","Buffer","from","arrayBuffer","staticProvider","model","tokenizerJson","dir","replace","recursive","file","console","error","headerLen","Number","readBigUInt64LE","header","JSON","parse","subarray","toString","entry","Object","entries","find","k","dtype","spec","dims","shape","dataStart","byteOffset","data_offsets","matrix","Float32Array","buffer","Tokenizer","tok","unkId","vocab","unk_token","one","ids","encode","add_special_tokens","id","isInteger","v","off","d","norm","Math","sqrt","texts","map","apiProvider","keyEnv","base","headers","key","process","env","undefined","authorization","post","method","body","stringify","input","json","data","embedding","providers","Map","getProvider","cfg","e","sig","type","p","get","set","toStore","full","int8","scale","max","abs","embedPending","baseDir","provider","dirty","all","storeDims","min","byPath","row","list","chunk","jobs","chunkIdxs","relPath","absPath","mtimeMs","size","doc","update","batch","vectors","j","job","q","Int8Array","round","err","semanticCandidates","terms","match","t","includes","qv","rows","best","vector","byteLength","dot","score","existing","start_line","end_line","sort","a","b","similarity"],"mappings":"AAAA,SAASA,UAAU,EAAEC,SAAS,EAAEC,YAAY,EAAEC,UAAU,EAAEC,aAAa,QAAQ,UAAU;AACzF,SAASC,OAAO,QAAQ,UAAU;AAClC,SAASC,IAAI,QAAQ,YAAY;AAGjC,SAASC,WAAW,QAAQ,eAAe;AAC3C,SAASC,UAAU,QAAQ,eAAe;AAC1C,SAASC,SAAS,QAAQ,aAAa;AAGvC,sFAAsF;AACtF,uFAAuF;AACvF,oFAAoF;AACpF,+EAA+E;AAE/E,6EAA6E;AAC7E,yEAAyE;AACzE,MAAMC,aAAa;AACnB,MAAMC,QAAQ;AAcd,uFAAuF;AACvF,sFAAsF;AACtF,mFAAmF;AACnF,yBAAyB;AACzB,SAASC,SAASC,GAAW,EAAEC,MAA2C;IACxE,MAAMC,QAAQF,IAAIG,KAAK,CAAC;IACxB,MAAMC,SAAmB,EAAE;IAC3B,IAAIC,UAAU;IACd,IAAK,IAAIC,IAAI,GAAGA,IAAIJ,MAAMK,MAAM,EAAED,IAAK;QACrC,IAAI,aAAaE,IAAI,CAACN,KAAK,CAACI,EAAE,GAAG;YAC/BD,UAAU,CAACA;YACX;QACF;QACA,IAAI,CAACA,WAAW,YAAYG,IAAI,CAACN,KAAK,CAACI,EAAE,GAAGF,OAAOK,IAAI,CAACH,IAAI;IAC9D;IACA,MAAMI,SAASN,OAAOG,MAAM,KAAK,IAAI;QAAC;KAAE,GAAGH,MAAM,CAAC,EAAE,GAAG,IAAI;QAAC;WAAMA;KAAO,GAAGA;IAC5E,MAAMO,SAAS;QAACV,mBAAAA,6BAAAA,OAAQW,KAAK;QAAEX,mBAAAA,6BAAAA,OAAQY,OAAO;KAAC,CAACC,MAAM,CAACC,SAAStB,IAAI,CAAC;IACrE,MAAMuB,SAAkB,EAAE;IAC1BN,OAAOO,OAAO,CAAC,CAACC,OAAOZ;QACrB,MAAMa,MAAMb,IAAI,IAAII,OAAOH,MAAM,GAAGG,MAAM,CAACJ,IAAI,EAAE,GAAG,IAAIJ,MAAMK,MAAM;QACpE,MAAMa,OAAOlB,MACVmB,KAAK,CAACH,QAAQ,GAAGC,KACjB1B,IAAI,CAAC,MACL6B,IAAI;QACP,IAAIF,KAAKb,MAAM,GAAG,GAAGS,OAAOP,IAAI,CAAC;YAAEc,WAAWL;YAAOM,SAASL;YAAKC,MAAMT,SAAS,GAAGA,OAAO,EAAE,EAAES,MAAM,GAAGA;QAAK;IAChH;IACA,OAAOJ;AACT;AAEA,OAAO,MAAMS,QAAiB;IAC5BC,MAAM;IACNC,QAAOC,EAAE;QACPA,GAAGC,IAAI,CAAC,CAAC,8JAA8J,CAAC;IAC1K;IACAC,SAAQ9B,GAAG,EAAE+B,KAAK,EAAE9B,MAAM;QACxB,OAAOF,SAASC,KAAKC;IACvB;IACA+B,QAAOJ,EAAE,EAAEK,IAAI;QACbL,GAAGM,OAAO,CAAC,2CAA2CC,GAAG,CAACF;IAC5D;IACAG,OAAMR,EAAE,EAAEK,IAAI,EAAEI,SAAS;QACvB,MAAMC,SAASV,GAAGM,OAAO,CAAC;QACzBG,UAAsBpB,OAAO,CAAC,CAACsB,GAAGC,MAAQF,OAAOH,GAAG,CAACF,MAAMO,KAAKD,EAAEhB,SAAS,EAAEgB,EAAEf,OAAO;IACzF;AACF,EAAE;AAEF,2FAA2F;AAC3F,yFAAyF;AACzF,gBAAgB;AAEhB,SAASiB,YAAYC,GAAW,EAAEC,IAAY;IAC5C,OAAOC,MAAMF,KAAKG,IAAI,CAAC,OAAOC;QAC5B,IAAI,CAACA,IAAIC,EAAE,EAAE,MAAM,IAAIpD,WAAW,eAAe,CAAC,uBAAuB,EAAE+C,IAAI,SAAS,EAAEI,IAAIE,MAAM,EAAE;QACtGzD,cAAc,GAAGoD,KAAK,KAAK,CAAC,EAAEM,OAAOC,IAAI,CAAC,MAAMJ,IAAIK,WAAW;QAC/D7D,WAAW,GAAGqD,KAAK,KAAK,CAAC,EAAEA;IAC7B;AACF;AAEA,eAAeS,eAAeC,KAAa;;QA2BEC,sBAA7BA,4BAAAA;IA1Bd,IAAIC,MAAMF;IACV,IAAI,CAAClE,WAAWM,KAAK8D,KAAK,uBAAuB;QAC/CA,MAAM9D,KAAKD,WAAW,UAAU,eAAe,UAAU6D,MAAMG,OAAO,CAAC,OAAO;QAC9EpE,UAAUmE,KAAK;YAAEE,WAAW;QAAK;QACjC,KAAK,MAAMC,QAAQ;YAAC;YAAqB;SAAiB,CAAE;YAC1D,IAAI,CAACvE,WAAWM,KAAK8D,KAAKG,QAAQ;gBAChCC,QAAQC,KAAK,CAAC,CAAC,SAAS,EAAEP,MAAM,CAAC,EAAEK,KAAK,MAAM,EAAEH,IAAI,0BAA0B,CAAC;gBAC/E,MAAMd,YAAY,CAAC,uBAAuB,EAAEY,MAAM,cAAc,EAAEK,MAAM,EAAEjE,KAAK8D,KAAKG;YACtF;QACF;IACF;IAEA,MAAM1D,MAAMX,aAAaI,KAAK8D,KAAK;IACnC,MAAMM,YAAYC,OAAO9D,IAAI+D,eAAe,CAAC;IAC7C,MAAMC,SAASC,KAAKC,KAAK,CAAClE,IAAImE,QAAQ,CAAC,GAAG,IAAIN,WAAWO,QAAQ,CAAC;IAClE,MAAMC,QAAQC,OAAOC,OAAO,CAACP,QAAQQ,IAAI,CAAC,CAAC,CAACC,EAAE,GAAKA,MAAM;IACzD,IAAI,CAACJ,SAASA,KAAK,CAAC,EAAE,CAACK,KAAK,KAAK,OAAO,MAAM,IAAI/E,WAAW,eAAe,GAAG0D,MAAM,oCAAoC,CAAC;IAC1H,MAAMsB,OAAON,KAAK,CAAC,EAAE;IACrB,MAAMO,OAAOD,KAAKE,KAAK,CAAC,EAAE;IAC1B,MAAMC,YAAY9E,IAAI+E,UAAU,GAAG,IAAIlB,YAAYc,KAAKK,YAAY,CAAC,EAAE;IACvE,MAAMC,SAASH,YAAY,MAAM,IAAI,IAAII,aAAalF,IAAImF,MAAM,EAAEL,WAAWH,KAAKE,KAAK,CAAC,EAAE,GAAGD,QAAQ,IAAIM,aAAalF,IAAImF,MAAM,CAAC9D,KAAK,CAACyD,WAAWA,YAAYH,KAAKE,KAAK,CAAC,EAAE,GAAGD,OAAO;IAErL,MAAMtB,gBAAgBW,KAAKC,KAAK,CAAC7E,aAAaI,KAAK8D,KAAK,mBAAmB;IAC3E,oFAAoF;IACpF,MAAM,EAAE6B,SAAS,EAAE,GAAG,MAAM,MAAM,CAAC;IACnC,MAAMC,MAAM,IAAID,UAAU9B,eAAe,CAAC;IAC1C,MAAMgC,iBAAQhC,wBAAAA,cAAcD,KAAK,cAAnBC,6CAAAA,6BAAAA,sBAAqBiC,KAAK,cAA1BjC,iDAAAA,0BAA4B,EAACA,uBAAAA,cAAcD,KAAK,cAAnBC,2CAAAA,qBAAqBkC,SAAS,CAAC,uCAAI,CAAC;IAE/E,SAASC,IAAIrE,IAAY;QACvB,mFAAmF;QACnF,uFAAuF;QACvF,yFAAyF;QACzF,MAAMsE,MAAM,AAACL,IAAIM,MAAM,CAACvE,MAAM;YAAEwE,oBAAoB;QAAM,GAAGF,GAAG,CAAc5E,MAAM,CAAC,CAAC+E,KAAO/B,OAAOgC,SAAS,CAACD,OAAOA,OAAOP;QAC5H,MAAMS,IAAI,IAAIb,aAAaN;QAC3B,IAAIc,IAAInF,MAAM,KAAK,GAAG,OAAOwF;QAC7B,KAAK,MAAMF,MAAMH,IAAK;YACpB,MAAMM,MAAMH,KAAKjB;YACjB,IAAK,IAAIqB,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKF,CAAC,CAACE,EAAE,IAAIhB,MAAM,CAACe,MAAMC,EAAE;QACxD;QACA,IAAIC,OAAO;QACX,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAK;YAC7BF,CAAC,CAACE,EAAE,IAAIP,IAAInF,MAAM;YAClB2F,QAAQH,CAAC,CAACE,EAAE,GAAGF,CAAC,CAACE,EAAE;QACrB;QACAC,OAAOC,KAAKC,IAAI,CAACF,QAAQ;QACzB,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKF,CAAC,CAACE,EAAE,IAAIC;QACvC,OAAOH;IACT;IAEA,OAAO;QAAEF,IAAI,CAAC,OAAO,EAAExC,OAAO;QAAEuB;QAAMnD,OAAO,OAAO4E,QAAUA,MAAMC,GAAG,CAACb;IAAK;AAC/E;AAEA,gFAAgF;AAEhF,eAAec,YAAYlD,KAAa,EAAEX,GAAuB,EAAE8D,MAA0B;IAC3F,IAAI,CAAC9D,KAAK,MAAM,IAAI/C,WAAW,eAAe;IAC9C,MAAM8G,OAAO/D,IAAIc,OAAO,CAAC,QAAQ;IACjC,MAAMkD,UAAkC;QAAE,gBAAgB;IAAmB;IAC7E,MAAMC,MAAMH,SAASI,QAAQC,GAAG,CAACL,OAAO,GAAGM;IAC3C,IAAIH,KAAKD,QAAQK,aAAa,GAAG,CAAC,OAAO,EAAEJ,KAAK;IAEhD,eAAeK,KAAKX,KAAe;QACjC,MAAMvD,MAAM,MAAMF,MAAM,GAAG6D,KAAK,WAAW,CAAC,EAAE;YAAEQ,QAAQ;YAAQP;YAASQ,MAAMjD,KAAKkD,SAAS,CAAC;gBAAE9D;gBAAO+D,OAAOf;YAAM;QAAG;QACvH,IAAI,CAACvD,IAAIC,EAAE,EAAE,MAAM,IAAIpD,WAAW,eAAe,GAAG8G,KAAK,oBAAoB,EAAE3D,IAAIE,MAAM,EAAE;QAC3F,MAAMkE,OAAQ,MAAMpE,IAAIuE,IAAI;QAC5B,OAAOH,KAAKI,IAAI,CAAChB,GAAG,CAAC,CAACL,IAAMf,aAAahC,IAAI,CAAC+C,EAAEsB,SAAS;IAC3D;IAEA,MAAM3C,OAAO,AAAC,CAAA,MAAMoC,KAAK;QAAC;KAAkB,CAAA,CAAE,CAAC,EAAE,CAACzG,MAAM;IACxD,OAAO;QAAEsF,IAAI,CAAC,IAAI,EAAEY,KAAK,CAAC,EAAEpD,OAAO;QAAEuB;QAAMnD,OAAOuF;IAAK;AACzD;AAEA,MAAMQ,YAAY,IAAIC;AAEtB,SAASC,YAAYC,GAAW;QAGMC;IAFpC,MAAMA,IAAIlI,YAAYiI;IACtB,IAAI,CAACC,GAAG,MAAM,IAAIjI,WAAW,kBAAkB;IAC/C,MAAMkI,MAAM,GAAGD,EAAEE,IAAI,CAAC,CAAC,EAAEF,EAAEvE,KAAK,CAAC,CAAC,GAAEuE,SAAAA,EAAElF,GAAG,cAALkF,oBAAAA,SAAS,IAAI;IACjD,IAAIG,IAAIP,UAAUQ,GAAG,CAACH;IACtB,IAAI,CAACE,GAAG;QACNA,IAAIH,EAAEE,IAAI,KAAK,QAAQvB,YAAYqB,EAAEvE,KAAK,EAAEuE,EAAElF,GAAG,EAAEkF,EAAEjB,GAAG,IAAIvD,eAAewE,EAAEvE,KAAK;QAClFmE,UAAUS,GAAG,CAACJ,KAAKE;IACrB;IACA,OAAOA;AACT;AAEA,4EAA4E;AAC5E,SAASG,QAAQC,IAAkB,EAAEvD,IAAY,EAAEwD,IAAa;IAC9D,MAAMrC,IAAI,IAAIb,aAAaN;IAC3B,IAAIsB,OAAO;IACX,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKC,QAAQiC,IAAI,CAAClC,EAAE,GAAGkC,IAAI,CAAClC,EAAE;IACxDC,OAAOC,KAAKC,IAAI,CAACF,QAAQ;IACzB,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKF,CAAC,CAACE,EAAE,GAAGkC,IAAI,CAAClC,EAAE,GAAGC;IAChD,IAAI,CAACkC,MAAM,OAAO;QAAErC;QAAGsC,OAAO;IAAE;IAChC,IAAIC,MAAM;IACV,IAAK,IAAIrC,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKqC,MAAMnC,KAAKmC,GAAG,CAACA,KAAKnC,KAAKoC,GAAG,CAACxC,CAAC,CAACE,EAAE;IAChE,OAAO;QAAEF;QAAGsC,OAAOC,MAAM,OAAO;IAAE;AACpC;AAEA,qFAAqF;AACrF,6CAA6C;AAC7C,OAAO,eAAeE,aAAa5G,EAAgB,EAAE+F,GAAW,EAAEc,OAAe;IAC/E,MAAMC,WAAW,MAAMhB,YAAYC,MAAM,kDAAkD;IAC3F,MAAMgB,QAAQ/G,GAAGM,OAAO,CAAC,oFAAoF0G,GAAG;IAChH,IAAID,MAAMpI,MAAM,KAAK,GAAG;IACxB,MAAMsI,YAAY1C,KAAK2C,GAAG,CAACjJ,YAAY6I,SAAS9D,IAAI;IAEpD,MAAMmE,SAAS,IAAItB;IACnB,KAAK,MAAMuB,OAAOL,MAAO;YACVI;QAAb,MAAME,QAAOF,cAAAA,OAAOf,GAAG,CAACgB,IAAI/G,IAAI,eAAnB8G,yBAAAA,cAAwB,EAAE;QACvCE,KAAKxI,IAAI,CAACuI,IAAIE,KAAK;QACnBH,OAAOd,GAAG,CAACe,IAAI/G,IAAI,EAAEgH;IACvB;IAEA,MAAME,OAA6D,EAAE;IACrE,KAAK,MAAM,CAAClH,MAAMmH,UAAU,IAAIL,OAAQ;QACtC,IAAI/H;QACJ,IAAI;YACFA,SAASpB,UAAU;gBAAEyJ,SAASpH;gBAAMqH,SAAS7J,KAAKgJ,SAASxG;gBAAOsH,SAAS;gBAAGC,MAAM;YAAE,GAAG;gBAAC/H;aAAM,EAAEgI,GAAG,CAACpH,SAAS,CAACZ,KAAK;QACvH,EAAE,OAAM;YACN,UAAU,gEAAgE;QAC5E;QACA,KAAK,MAAMe,OAAO4G,UAAW,IAAIpI,MAAM,CAACwB,IAAI,EAAE2G,KAAK1I,IAAI,CAAC;YAAEwB;YAAMiH,OAAO1G;YAAKpB,MAAMJ,MAAM,CAACwB,IAAI,CAACpB,IAAI;QAAC;IACrG;IAEA,MAAMsI,SAAS9H,GAAGM,OAAO,CAAC;IAC1B,IAAK,IAAI5B,IAAI,GAAGA,IAAI6I,KAAK5I,MAAM,EAAED,KAAKR,MAAO;QAC3C,MAAM6J,QAAQR,KAAK9H,KAAK,CAACf,GAAGA,IAAIR;QAChC,MAAM8J,UAAU,MAAMlB,SAASjH,KAAK,CAACkI,MAAMrD,GAAG,CAAC,CAACuD,IAAMA,EAAEzI,IAAI;QAC5DQ,GAAGC,IAAI,CAAC;QACR,IAAI;YACF8H,MAAM1I,OAAO,CAAC,CAAC6I,KAAKD;gBAClB,MAAM,EAAE9D,CAAC,EAAEsC,KAAK,EAAE,GAAGH,QAAQ0B,OAAO,CAACC,EAAE,EAAEhB,WAAW;gBACpD,MAAMkB,IAAI,IAAIC,UAAUnB;gBACxB,IAAK,IAAI5C,IAAI,GAAGA,IAAI4C,WAAW5C,IAAK8D,CAAC,CAAC9D,EAAE,GAAGE,KAAK8D,KAAK,CAAClE,CAAC,CAACE,EAAE,GAAGoC;gBAC7DqB,OAAOvH,GAAG,CAACkG,OAAOpF,OAAOC,IAAI,CAAC6G,EAAE5E,MAAM,GAAG2E,IAAI7H,IAAI,EAAE6H,IAAIZ,KAAK;YAC9D;YACAtH,GAAGC,IAAI,CAAC;QACV,EAAE,OAAOqI,KAAK;YACZtI,GAAGC,IAAI,CAAC;YACR,MAAMqI;QACR;IACF;AACF;AAEA,oFAAoF;AACpF,wFAAwF;AACxF,oBAAoB;AACpB,4FAA4F;AAC5F,8FAA8F;AAC9F,2FAA2F;AAC3F,6FAA6F;AAC7F,8BAA8B;AAC9B,OAAO,eAAeC,mBAAmBvI,EAAgB,EAAE+F,GAAW,EAAEyC,KAAa,EAAExH,MAAa;QAOpFwH;IANd,MAAM3B,UAAU,AAACd,IAAgCc,OAAO;IACxD,IAAI,CAACA,SAAS,MAAM,IAAI9I,WAAW,eAAe;IAClD,MAAM6I,aAAa5G,IAAI+F,KAAKc;IAE5B,MAAMC,WAAW,MAAMhB,YAAYC;IACnC,MAAMkB,YAAY1C,KAAK2C,GAAG,CAACjJ,YAAY6I,SAAS9D,IAAI;IACpD,MAAMxD,OAAO,EAACgJ,eAAAA,MAAMC,KAAK,CAAC,gCAAZD,0BAAAA,eAAkC,EAAE,EAAEtJ,MAAM,CAAC,CAACwJ,IAAM,CAAC;YAAC;YAAO;YAAM;YAAO;SAAO,CAACC,QAAQ,CAACD,IAAI7K,IAAI,CAAC;IAClH,MAAM,EAAEsG,GAAGyE,EAAE,EAAE,GAAGtC,QAAQ,AAAC,CAAA,MAAMQ,SAASjH,KAAK,CAAC;QAACL;KAAK,CAAA,CAAE,CAAC,EAAE,EAAEyH,WAAW;IAExE,MAAM4B,OAAO7I,GAAGM,OAAO,CAAC,+FAA+F0G,GAAG;IAQ1H,MAAM8B,OAAO,IAAIjD;IACjB,KAAK,MAAMuB,OAAOyB,KAAM;QACtB,MAAMV,IAAI,IAAIC,UAAUhB,IAAI2B,MAAM,CAACxF,MAAM,EAAE6D,IAAI2B,MAAM,CAAC5F,UAAU,EAAEoB,KAAK2C,GAAG,CAACD,WAAWG,IAAI2B,MAAM,CAACC,UAAU;QAC3G,IAAIC,MAAM;QACV,IAAK,IAAI5E,IAAI,GAAGA,IAAI8D,EAAExJ,MAAM,EAAE0F,IAAK4E,OAAOd,CAAC,CAAC9D,EAAE,GAAGuE,EAAE,CAACvE,EAAE;QACtD,MAAM6E,QAAQD,MAAM7B,IAAIX,KAAK;QAC7B,MAAM0C,WAAWL,KAAK1C,GAAG,CAACgB,IAAI/G,IAAI;QAClC,IAAI,CAAC8I,YAAYD,QAAQC,SAASD,KAAK,EAAEJ,KAAKzC,GAAG,CAACe,IAAI/G,IAAI,EAAE;YAAE6I;YAAO5K,OAAO,CAAC,CAAC,EAAE8I,IAAIgC,UAAU,CAAC,CAAC,EAAEhC,IAAIiC,QAAQ,EAAE;QAAC;IACnH;IACA,OAAO;WAAIP,KAAKnG,OAAO;KAAG,CACvB2G,IAAI,CAAC,CAACC,GAAGC,IAAMA,CAAC,CAAC,EAAE,CAACN,KAAK,GAAGK,CAAC,CAAC,EAAE,CAACL,KAAK,EACtCzJ,KAAK,CAAC,GAAGuB,QACT0D,GAAG,CAAC,CAAC,CAACrE,MAAMmJ,EAAE,GAAM,CAAA;YAAEnJ;YAAM/B,OAAOkL,EAAElL,KAAK;YAAEmL,YAAYlF,KAAK8D,KAAK,CAACmB,EAAEN,KAAK,GAAG,QAAQ;QAAK,CAAA;AAC/F"}
+{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/embed.ts"],"sourcesContent":["import { existsSync, mkdirSync, readFileSync, renameSync, writeFileSync } from 'node:fs';\nimport { homedir } from 'node:os';\nimport { join } from 'node:path';\nimport type { DatabaseSync } from 'node:sqlite';\nimport type { Config, ResolvedConfig } from '../config.ts';\nimport { embedConfig } from '../config.ts';\nimport { SenseError } from '../errors.ts';\nimport { progress } from '../progress.ts';\nimport { parseFile } from '../scan.ts';\nimport type { Feature } from './types.ts';\n\n// embeddings(path, chunk, start_line, end_line, scale, vector): heading-based chunks,\n// int8 vectors with a per-vector dequantization scale, vector NULL = not yet embedded.\n// Reconcile stores dirty rows inside its transaction; embedding tops up on the next\n// semantic query (embedPending), so staleness costs recall, never correctness.\n\n// Storage lever fixed by the bake-off (BENCHMARKING.md): int8 at 256 dims is\n// quality-free vs f32-512 when fused. Queries stay f32 at the same dims.\nconst STORE_DIMS = 256;\nconst BATCH = 64;\n\nexport interface EmbedProvider {\n  id: string; // model identity; participates in the cache key, change -> re-embed\n  dims: number;\n  embed(texts: string[]): Promise<Float32Array[]>;\n}\n\ninterface Chunk {\n  startLine: number;\n  endLine: number;\n  text: string;\n}\n\n// Deterministic chunker used at reconcile (line ranges stored) and at embed time (text\n// re-derived from the file -- chunk text is never stored). Heading-delimited with the\n// preamble kept; whole file when no headings. The title/summary prefix mirrors the\n// bm25 column weighting.\nfunction chunksOf(raw: string, search?: { title: string; summary: string }): Chunk[] {\n  const lines = raw.split('\\n');\n  const starts: number[] = [];\n  let inFence = false;\n  for (let i = 0; i < lines.length; i++) {\n    if (/^(```|~~~)/.test(lines[i])) {\n      inFence = !inFence;\n      continue;\n    }\n    if (!inFence && /^#{1,6} +/.test(lines[i])) starts.push(i + 1);\n  }\n  const bounds = starts.length === 0 ? [1] : starts[0] > 1 ? [1, ...starts] : starts;\n  const prefix = [search?.title, search?.summary].filter(Boolean).join('\\n');\n  const chunks: Chunk[] = [];\n  bounds.forEach((start, i) => {\n    const end = i + 1 < bounds.length ? bounds[i + 1] - 1 : lines.length;\n    const text = lines\n      .slice(start - 1, end)\n      .join('\\n')\n      .trim();\n    if (text.length > 0) chunks.push({ startLine: start, endLine: end, text: prefix ? `${prefix}\\n${text}` : text });\n  });\n  return chunks;\n}\n\nexport const embed: Feature = {\n  name: 'embed',\n  schema(db) {\n    db.exec(`CREATE TABLE IF NOT EXISTS embeddings (\"path\" TEXT, chunk INTEGER, start_line INTEGER, end_line INTEGER, scale REAL, vector BLOB, PRIMARY KEY (\"path\", chunk))`);\n  },\n  extract(raw, _body, search) {\n    return chunksOf(raw, search);\n  },\n  remove(db, path) {\n    db.prepare('DELETE FROM embeddings WHERE \"path\" = ?').run(path);\n  },\n  // A doc covered only by semantic-false presets never had extract() run for it (db.ts's\n  // per-file filter skips it), so extracted is undefined here -- store nothing, i.e. no rows.\n  store(db, path, extracted) {\n    if (!extracted) return;\n    const insert = db.prepare('INSERT INTO embeddings (\"path\", chunk, start_line, end_line, scale, vector) VALUES (?, ?, ?, ?, NULL, NULL)');\n    (extracted as Chunk[]).forEach((c, idx) => insert.run(path, idx, c.startLine, c.endLine));\n  },\n  enabledForFile(_cfg, file) {\n    return file.embed;\n  },\n};\n\n// --- static type: Model2Vec safetensors + pure-JS tokenizer, model cached in ~/.cache ---\n// Encode convention from model2vec/model.py: no special tokens, drop unk ids, mean-pool,\n// L2-normalize.\n\nfunction fetchToFile(url: string, dest: string): Promise<void> {\n  return fetch(url).then(async (res) => {\n    if (!res.ok) throw new SenseError('EMBED_MODEL', `model download failed: ${url} -> HTTP ${res.status}`);\n    writeFileSync(`${dest}.part`, Buffer.from(await res.arrayBuffer()));\n    renameSync(`${dest}.part`, dest);\n  });\n}\n\nasync function staticProvider(model: string): Promise<EmbedProvider> {\n  let dir = model;\n  if (!existsSync(join(dir, 'model.safetensors'))) {\n    dir = join(homedir(), '.cache', 'sensemaking', 'models', model.replace(/\\//g, '--'));\n    mkdirSync(dir, { recursive: true });\n    for (const file of ['model.safetensors', 'tokenizer.json']) {\n      if (!existsSync(join(dir, file))) {\n        console.error(`fetching ${model}/${file} into ${dir} (once; delete to refetch)`);\n        await fetchToFile(`https://huggingface.co/${model}/resolve/main/${file}`, join(dir, file));\n      }\n    }\n  }\n\n  const raw = readFileSync(join(dir, 'model.safetensors'));\n  const headerLen = Number(raw.readBigUInt64LE(0));\n  const header = JSON.parse(raw.subarray(8, 8 + headerLen).toString('utf8')) as Record<string, { dtype: string; shape: number[]; data_offsets: number[] }>;\n  const entry = Object.entries(header).find(([k]) => k !== '__metadata__');\n  if (!entry || entry[1].dtype !== 'F32') throw new SenseError('EMBED_MODEL', `${model}: expected an F32 safetensors matrix`);\n  const spec = entry[1];\n  const dims = spec.shape[1];\n  const dataStart = raw.byteOffset + 8 + headerLen + spec.data_offsets[0];\n  const matrix = dataStart % 4 === 0 ? new Float32Array(raw.buffer, dataStart, spec.shape[0] * dims) : new Float32Array(raw.buffer.slice(dataStart, dataStart + spec.shape[0] * dims * 4));\n\n  const tokenizerJson = JSON.parse(readFileSync(join(dir, 'tokenizer.json'), 'utf8'));\n  // Lazy import: the tokenizer loads only on the semantic path, never at CLI startup.\n  const { Tokenizer } = await import('@huggingface/tokenizers');\n  const tok = new Tokenizer(tokenizerJson, {});\n  const unkId = tokenizerJson.model?.vocab?.[tokenizerJson.model?.unk_token] ?? -1;\n\n  function one(text: string): Float32Array {\n    // The tokenizer yields undefined (not the unk id) for tokens outside the vocab; an\n    // undefined id would index the matrix at NaN and poison the whole mean-pool -- and the\n    // int8 conversion then stores the NaN vector as all zeros, silently. Keep integers only.\n    const ids = (tok.encode(text, { add_special_tokens: false }).ids as number[]).filter((id) => Number.isInteger(id) && id !== unkId);\n    const v = new Float32Array(dims);\n    if (ids.length === 0) return v;\n    for (const id of ids) {\n      const off = id * dims;\n      for (let d = 0; d < dims; d++) v[d] += matrix[off + d];\n    }\n    let norm = 0;\n    for (let d = 0; d < dims; d++) {\n      v[d] /= ids.length;\n      norm += v[d] * v[d];\n    }\n    norm = Math.sqrt(norm) + 1e-32;\n    for (let d = 0; d < dims; d++) v[d] /= norm;\n    return v;\n  }\n\n  return { id: `static:${model}`, dims, embed: async (texts) => texts.map(one) };\n}\n\n// --- api type: one POST against any OpenAI-compatible /embeddings endpoint ---\n\nasync function apiProvider(model: string, url: string | undefined, keyEnv: string | undefined): Promise<EmbedProvider> {\n  if (!url) throw new SenseError('EMBED_MODEL', 'features.embed.type \"api\" requires a url');\n  const base = url.replace(/\\/+$/, '');\n  const headers: Record<string, string> = { 'content-type': 'application/json' };\n  const key = keyEnv ? process.env[keyEnv] : undefined;\n  if (key) headers.authorization = `Bearer ${key}`;\n\n  async function post(texts: string[]): Promise<Float32Array[]> {\n    const res = await fetch(`${base}/embeddings`, { method: 'POST', headers, body: JSON.stringify({ model, input: texts }) });\n    if (!res.ok) throw new SenseError('EMBED_MODEL', `${base}/embeddings -> HTTP ${res.status}`);\n    const body = (await res.json()) as { data: Array<{ embedding: number[] }> };\n    return body.data.map((d) => Float32Array.from(d.embedding));\n  }\n\n  const dims = (await post(['dimension probe']))[0].length;\n  return { id: `api:${base}:${model}`, dims, embed: post };\n}\n\nconst providers = new Map<string, Promise<EmbedProvider>>();\n\nfunction getProvider(cfg: Config): Promise<EmbedProvider> {\n  const e = embedConfig(cfg);\n  if (!e) throw new SenseError('EMBED_DISABLED', 'semantic search needs at least one preset with vectors on (semantic !== false); every declared preset in sense.config.json has \"semantic\": false');\n  const sig = `${e.type}:${e.model}:${e.url ?? ''}`;\n  let p = providers.get(sig);\n  if (!p) {\n    p = e.type === 'api' ? apiProvider(e.model, e.url, e.key) : staticProvider(e.model);\n    providers.set(sig, p);\n  }\n  return p;\n}\n\n// Slice + re-normalize (Matryoshka); optionally round through int8 storage.\nfunction toStore(full: Float32Array, dims: number, int8: boolean): { v: Float32Array; scale: number } {\n  const v = new Float32Array(dims);\n  let norm = 0;\n  for (let d = 0; d < dims; d++) norm += full[d] * full[d];\n  norm = Math.sqrt(norm) + 1e-32;\n  for (let d = 0; d < dims; d++) v[d] = full[d] / norm;\n  if (!int8) return { v, scale: 1 };\n  let max = 0;\n  for (let d = 0; d < dims; d++) max = Math.max(max, Math.abs(v[d]));\n  return { v, scale: max / 127 || 1 };\n}\n\n// Embed rows whose vector is NULL, re-deriving chunk text from the files through the\n// same parse + chunker that stored the rows.\nexport async function embedPending(db: DatabaseSync, cfg: Config, baseDir: string): Promise<void> {\n  const provider = await getProvider(cfg); // throws EMBED_DISABLED before touching the table\n  const dirty = db.prepare('SELECT \"path\", chunk FROM embeddings WHERE vector IS NULL ORDER BY \"path\", chunk').all() as Array<{ path: string; chunk: number }>;\n  if (dirty.length === 0) return;\n  const storeDims = Math.min(STORE_DIMS, provider.dims);\n\n  const byPath = new Map<string, number[]>();\n  for (const row of dirty) {\n    const list = byPath.get(row.path) ?? [];\n    list.push(row.chunk);\n    byPath.set(row.path, list);\n  }\n\n  const jobs: Array<{ path: string; chunk: number; text: string }> = [];\n  for (const [path, chunkIdxs] of byPath) {\n    let chunks: Chunk[];\n    try {\n      // presets/embed are irrelevant here -- re-deriving chunk text for a doc that already\n      // has embeddings rows means it was covered by a semantic-on preset at reconcile time.\n      chunks = parseFile({ relPath: path, absPath: join(baseDir, path), mtimeMs: 0, size: 0, presets: [], embed: true }, [embed]).doc.extracted.embed as Chunk[];\n    } catch {\n      continue; // vanished since reconcile; the next reconcile removes its rows\n    }\n    for (const idx of chunkIdxs) if (chunks[idx]) jobs.push({ path, chunk: idx, text: chunks[idx].text });\n  }\n\n  const update = db.prepare('UPDATE embeddings SET scale = ?, vector = ? WHERE \"path\" = ? AND chunk = ?');\n  // The lazy build is the one long silence a first --semantic query hits (measured 23s at\n  // 26k notes); progress makes it distinguishable from a hang.\n  const report = progress('embedding chunks', jobs.length);\n  for (let i = 0; i < jobs.length; i += BATCH) {\n    const batch = jobs.slice(i, i + BATCH);\n    const vectors = await provider.embed(batch.map((j) => j.text));\n    db.exec('BEGIN');\n    try {\n      batch.forEach((job, j) => {\n        const { v, scale } = toStore(vectors[j], storeDims, true);\n        const q = new Int8Array(storeDims);\n        for (let d = 0; d < storeDims; d++) q[d] = Math.round(v[d] / scale);\n        update.run(scale, Buffer.from(q.buffer), job.path, job.chunk);\n      });\n      db.exec('COMMIT');\n    } catch (err) {\n      db.exec('ROLLBACK');\n      throw err;\n    }\n    report.tick(Math.min(i + BATCH, jobs.length));\n  }\n  report.finish();\n}\n\n// Top candidates by cosine for a semantic find: best chunk per file, its line range\n// riding along. FTS5 operators in the terms are lexical syntax, not meaning -- stripped\n// before embedding.\n// Returns the cosine similarity alongside each candidate: stored and query vectors are both\n// L2-normalised before quantisation, so the dot product is cosine in [-1, 1]. `find` surfaces\n// it because the fused rank score cannot express match quality -- nearest-neighbour search\n// always returns a nearest neighbour, so without a magnitude an agent cannot tell a real hit\n// from the best of a bad lot.\nexport async function semanticCandidates(db: DatabaseSync, cfg: Config, terms: string, fetch: number): Promise<Array<{ path: string; lines: string; similarity: number }>> {\n  const baseDir = (cfg as Partial<ResolvedConfig>).baseDir;\n  if (!baseDir) throw new SenseError('EMBED_MODEL', 'semantic expansion needs a config with baseDir (use loadConfig/open)');\n  await embedPending(db, cfg, baseDir);\n\n  const provider = await getProvider(cfg);\n  const storeDims = Math.min(STORE_DIMS, provider.dims);\n  const text = (terms.match(/[\\p{L}\\p{N}]+/gu) ?? []).filter((t) => !['AND', 'OR', 'NOT', 'NEAR'].includes(t)).join(' ');\n  const { v: qv } = toStore((await provider.embed([text]))[0], storeDims, false);\n\n  const rows = db.prepare('SELECT \"path\", start_line, end_line, scale, vector FROM embeddings WHERE vector IS NOT NULL').all() as Array<{\n    path: string;\n    start_line: number;\n    end_line: number;\n    scale: number;\n    vector: Uint8Array;\n  }>;\n\n  const best = new Map<string, { score: number; lines: string }>();\n  for (const row of rows) {\n    const q = new Int8Array(row.vector.buffer, row.vector.byteOffset, Math.min(storeDims, row.vector.byteLength));\n    let dot = 0;\n    for (let d = 0; d < q.length; d++) dot += q[d] * qv[d];\n    const score = dot * row.scale;\n    const existing = best.get(row.path);\n    if (!existing || score > existing.score) best.set(row.path, { score, lines: `L${row.start_line}-${row.end_line}` });\n  }\n  return [...best.entries()]\n    .sort((a, b) => b[1].score - a[1].score)\n    .slice(0, fetch)\n    .map(([path, b]) => ({ path, lines: b.lines, similarity: Math.round(b.score * 1000) / 1000 }));\n}\n"],"names":["existsSync","mkdirSync","readFileSync","renameSync","writeFileSync","homedir","join","embedConfig","SenseError","progress","parseFile","STORE_DIMS","BATCH","chunksOf","raw","search","lines","split","starts","inFence","i","length","test","push","bounds","prefix","title","summary","filter","Boolean","chunks","forEach","start","end","text","slice","trim","startLine","endLine","embed","name","schema","db","exec","extract","_body","remove","path","prepare","run","store","extracted","insert","c","idx","enabledForFile","_cfg","file","fetchToFile","url","dest","fetch","then","res","ok","status","Buffer","from","arrayBuffer","staticProvider","model","tokenizerJson","dir","replace","recursive","console","error","headerLen","Number","readBigUInt64LE","header","JSON","parse","subarray","toString","entry","Object","entries","find","k","dtype","spec","dims","shape","dataStart","byteOffset","data_offsets","matrix","Float32Array","buffer","Tokenizer","tok","unkId","vocab","unk_token","one","ids","encode","add_special_tokens","id","isInteger","v","off","d","norm","Math","sqrt","texts","map","apiProvider","keyEnv","base","headers","key","process","env","undefined","authorization","post","method","body","stringify","input","json","data","embedding","providers","Map","getProvider","cfg","e","sig","type","p","get","set","toStore","full","int8","scale","max","abs","embedPending","baseDir","provider","dirty","all","storeDims","min","byPath","row","list","chunk","jobs","chunkIdxs","relPath","absPath","mtimeMs","size","presets","doc","update","report","batch","vectors","j","job","q","Int8Array","round","err","tick","finish","semanticCandidates","terms","match","t","includes","qv","rows","best","vector","byteLength","dot","score","existing","start_line","end_line","sort","a","b","similarity"],"mappings":"AAAA,SAASA,UAAU,EAAEC,SAAS,EAAEC,YAAY,EAAEC,UAAU,EAAEC,aAAa,QAAQ,UAAU;AACzF,SAASC,OAAO,QAAQ,UAAU;AAClC,SAASC,IAAI,QAAQ,YAAY;AAGjC,SAASC,WAAW,QAAQ,eAAe;AAC3C,SAASC,UAAU,QAAQ,eAAe;AAC1C,SAASC,QAAQ,QAAQ,iBAAiB;AAC1C,SAASC,SAAS,QAAQ,aAAa;AAGvC,sFAAsF;AACtF,uFAAuF;AACvF,oFAAoF;AACpF,+EAA+E;AAE/E,6EAA6E;AAC7E,yEAAyE;AACzE,MAAMC,aAAa;AACnB,MAAMC,QAAQ;AAcd,uFAAuF;AACvF,sFAAsF;AACtF,mFAAmF;AACnF,yBAAyB;AACzB,SAASC,SAASC,GAAW,EAAEC,MAA2C;IACxE,MAAMC,QAAQF,IAAIG,KAAK,CAAC;IACxB,MAAMC,SAAmB,EAAE;IAC3B,IAAIC,UAAU;IACd,IAAK,IAAIC,IAAI,GAAGA,IAAIJ,MAAMK,MAAM,EAAED,IAAK;QACrC,IAAI,aAAaE,IAAI,CAACN,KAAK,CAACI,EAAE,GAAG;YAC/BD,UAAU,CAACA;YACX;QACF;QACA,IAAI,CAACA,WAAW,YAAYG,IAAI,CAACN,KAAK,CAACI,EAAE,GAAGF,OAAOK,IAAI,CAACH,IAAI;IAC9D;IACA,MAAMI,SAASN,OAAOG,MAAM,KAAK,IAAI;QAAC;KAAE,GAAGH,MAAM,CAAC,EAAE,GAAG,IAAI;QAAC;WAAMA;KAAO,GAAGA;IAC5E,MAAMO,SAAS;QAACV,mBAAAA,6BAAAA,OAAQW,KAAK;QAAEX,mBAAAA,6BAAAA,OAAQY,OAAO;KAAC,CAACC,MAAM,CAACC,SAASvB,IAAI,CAAC;IACrE,MAAMwB,SAAkB,EAAE;IAC1BN,OAAOO,OAAO,CAAC,CAACC,OAAOZ;QACrB,MAAMa,MAAMb,IAAI,IAAII,OAAOH,MAAM,GAAGG,MAAM,CAACJ,IAAI,EAAE,GAAG,IAAIJ,MAAMK,MAAM;QACpE,MAAMa,OAAOlB,MACVmB,KAAK,CAACH,QAAQ,GAAGC,KACjB3B,IAAI,CAAC,MACL8B,IAAI;QACP,IAAIF,KAAKb,MAAM,GAAG,GAAGS,OAAOP,IAAI,CAAC;YAAEc,WAAWL;YAAOM,SAASL;YAAKC,MAAMT,SAAS,GAAGA,OAAO,EAAE,EAAES,MAAM,GAAGA;QAAK;IAChH;IACA,OAAOJ;AACT;AAEA,OAAO,MAAMS,QAAiB;IAC5BC,MAAM;IACNC,QAAOC,EAAE;QACPA,GAAGC,IAAI,CAAC,CAAC,8JAA8J,CAAC;IAC1K;IACAC,SAAQ9B,GAAG,EAAE+B,KAAK,EAAE9B,MAAM;QACxB,OAAOF,SAASC,KAAKC;IACvB;IACA+B,QAAOJ,EAAE,EAAEK,IAAI;QACbL,GAAGM,OAAO,CAAC,2CAA2CC,GAAG,CAACF;IAC5D;IACA,uFAAuF;IACvF,4FAA4F;IAC5FG,OAAMR,EAAE,EAAEK,IAAI,EAAEI,SAAS;QACvB,IAAI,CAACA,WAAW;QAChB,MAAMC,SAASV,GAAGM,OAAO,CAAC;QACzBG,UAAsBpB,OAAO,CAAC,CAACsB,GAAGC,MAAQF,OAAOH,GAAG,CAACF,MAAMO,KAAKD,EAAEhB,SAAS,EAAEgB,EAAEf,OAAO;IACzF;IACAiB,gBAAeC,IAAI,EAAEC,IAAI;QACvB,OAAOA,KAAKlB,KAAK;IACnB;AACF,EAAE;AAEF,2FAA2F;AAC3F,yFAAyF;AACzF,gBAAgB;AAEhB,SAASmB,YAAYC,GAAW,EAAEC,IAAY;IAC5C,OAAOC,MAAMF,KAAKG,IAAI,CAAC,OAAOC;QAC5B,IAAI,CAACA,IAAIC,EAAE,EAAE,MAAM,IAAIxD,WAAW,eAAe,CAAC,uBAAuB,EAAEmD,IAAI,SAAS,EAAEI,IAAIE,MAAM,EAAE;QACtG7D,cAAc,GAAGwD,KAAK,KAAK,CAAC,EAAEM,OAAOC,IAAI,CAAC,MAAMJ,IAAIK,WAAW;QAC/DjE,WAAW,GAAGyD,KAAK,KAAK,CAAC,EAAEA;IAC7B;AACF;AAEA,eAAeS,eAAeC,KAAa;;QA2BEC,sBAA7BA,4BAAAA;IA1Bd,IAAIC,MAAMF;IACV,IAAI,CAACtE,WAAWM,KAAKkE,KAAK,uBAAuB;QAC/CA,MAAMlE,KAAKD,WAAW,UAAU,eAAe,UAAUiE,MAAMG,OAAO,CAAC,OAAO;QAC9ExE,UAAUuE,KAAK;YAAEE,WAAW;QAAK;QACjC,KAAK,MAAMjB,QAAQ;YAAC;YAAqB;SAAiB,CAAE;YAC1D,IAAI,CAACzD,WAAWM,KAAKkE,KAAKf,QAAQ;gBAChCkB,QAAQC,KAAK,CAAC,CAAC,SAAS,EAAEN,MAAM,CAAC,EAAEb,KAAK,MAAM,EAAEe,IAAI,0BAA0B,CAAC;gBAC/E,MAAMd,YAAY,CAAC,uBAAuB,EAAEY,MAAM,cAAc,EAAEb,MAAM,EAAEnD,KAAKkE,KAAKf;YACtF;QACF;IACF;IAEA,MAAM3C,MAAMZ,aAAaI,KAAKkE,KAAK;IACnC,MAAMK,YAAYC,OAAOhE,IAAIiE,eAAe,CAAC;IAC7C,MAAMC,SAASC,KAAKC,KAAK,CAACpE,IAAIqE,QAAQ,CAAC,GAAG,IAAIN,WAAWO,QAAQ,CAAC;IAClE,MAAMC,QAAQC,OAAOC,OAAO,CAACP,QAAQQ,IAAI,CAAC,CAAC,CAACC,EAAE,GAAKA,MAAM;IACzD,IAAI,CAACJ,SAASA,KAAK,CAAC,EAAE,CAACK,KAAK,KAAK,OAAO,MAAM,IAAIlF,WAAW,eAAe,GAAG8D,MAAM,oCAAoC,CAAC;IAC1H,MAAMqB,OAAON,KAAK,CAAC,EAAE;IACrB,MAAMO,OAAOD,KAAKE,KAAK,CAAC,EAAE;IAC1B,MAAMC,YAAYhF,IAAIiF,UAAU,GAAG,IAAIlB,YAAYc,KAAKK,YAAY,CAAC,EAAE;IACvE,MAAMC,SAASH,YAAY,MAAM,IAAI,IAAII,aAAapF,IAAIqF,MAAM,EAAEL,WAAWH,KAAKE,KAAK,CAAC,EAAE,GAAGD,QAAQ,IAAIM,aAAapF,IAAIqF,MAAM,CAAChE,KAAK,CAAC2D,WAAWA,YAAYH,KAAKE,KAAK,CAAC,EAAE,GAAGD,OAAO;IAErL,MAAMrB,gBAAgBU,KAAKC,KAAK,CAAChF,aAAaI,KAAKkE,KAAK,mBAAmB;IAC3E,oFAAoF;IACpF,MAAM,EAAE4B,SAAS,EAAE,GAAG,MAAM,MAAM,CAAC;IACnC,MAAMC,MAAM,IAAID,UAAU7B,eAAe,CAAC;IAC1C,MAAM+B,iBAAQ/B,wBAAAA,cAAcD,KAAK,cAAnBC,6CAAAA,6BAAAA,sBAAqBgC,KAAK,cAA1BhC,iDAAAA,0BAA4B,EAACA,uBAAAA,cAAcD,KAAK,cAAnBC,2CAAAA,qBAAqBiC,SAAS,CAAC,uCAAI,CAAC;IAE/E,SAASC,IAAIvE,IAAY;QACvB,mFAAmF;QACnF,uFAAuF;QACvF,yFAAyF;QACzF,MAAMwE,MAAM,AAACL,IAAIM,MAAM,CAACzE,MAAM;YAAE0E,oBAAoB;QAAM,GAAGF,GAAG,CAAc9E,MAAM,CAAC,CAACiF,KAAO/B,OAAOgC,SAAS,CAACD,OAAOA,OAAOP;QAC5H,MAAMS,IAAI,IAAIb,aAAaN;QAC3B,IAAIc,IAAIrF,MAAM,KAAK,GAAG,OAAO0F;QAC7B,KAAK,MAAMF,MAAMH,IAAK;YACpB,MAAMM,MAAMH,KAAKjB;YACjB,IAAK,IAAIqB,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKF,CAAC,CAACE,EAAE,IAAIhB,MAAM,CAACe,MAAMC,EAAE;QACxD;QACA,IAAIC,OAAO;QACX,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAK;YAC7BF,CAAC,CAACE,EAAE,IAAIP,IAAIrF,MAAM;YAClB6F,QAAQH,CAAC,CAACE,EAAE,GAAGF,CAAC,CAACE,EAAE;QACrB;QACAC,OAAOC,KAAKC,IAAI,CAACF,QAAQ;QACzB,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKF,CAAC,CAACE,EAAE,IAAIC;QACvC,OAAOH;IACT;IAEA,OAAO;QAAEF,IAAI,CAAC,OAAO,EAAEvC,OAAO;QAAEsB;QAAMrD,OAAO,OAAO8E,QAAUA,MAAMC,GAAG,CAACb;IAAK;AAC/E;AAEA,gFAAgF;AAEhF,eAAec,YAAYjD,KAAa,EAAEX,GAAuB,EAAE6D,MAA0B;IAC3F,IAAI,CAAC7D,KAAK,MAAM,IAAInD,WAAW,eAAe;IAC9C,MAAMiH,OAAO9D,IAAIc,OAAO,CAAC,QAAQ;IACjC,MAAMiD,UAAkC;QAAE,gBAAgB;IAAmB;IAC7E,MAAMC,MAAMH,SAASI,QAAQC,GAAG,CAACL,OAAO,GAAGM;IAC3C,IAAIH,KAAKD,QAAQK,aAAa,GAAG,CAAC,OAAO,EAAEJ,KAAK;IAEhD,eAAeK,KAAKX,KAAe;QACjC,MAAMtD,MAAM,MAAMF,MAAM,GAAG4D,KAAK,WAAW,CAAC,EAAE;YAAEQ,QAAQ;YAAQP;YAASQ,MAAMjD,KAAKkD,SAAS,CAAC;gBAAE7D;gBAAO8D,OAAOf;YAAM;QAAG;QACvH,IAAI,CAACtD,IAAIC,EAAE,EAAE,MAAM,IAAIxD,WAAW,eAAe,GAAGiH,KAAK,oBAAoB,EAAE1D,IAAIE,MAAM,EAAE;QAC3F,MAAMiE,OAAQ,MAAMnE,IAAIsE,IAAI;QAC5B,OAAOH,KAAKI,IAAI,CAAChB,GAAG,CAAC,CAACL,IAAMf,aAAa/B,IAAI,CAAC8C,EAAEsB,SAAS;IAC3D;IAEA,MAAM3C,OAAO,AAAC,CAAA,MAAMoC,KAAK;QAAC;KAAkB,CAAA,CAAE,CAAC,EAAE,CAAC3G,MAAM;IACxD,OAAO;QAAEwF,IAAI,CAAC,IAAI,EAAEY,KAAK,CAAC,EAAEnD,OAAO;QAAEsB;QAAMrD,OAAOyF;IAAK;AACzD;AAEA,MAAMQ,YAAY,IAAIC;AAEtB,SAASC,YAAYC,GAAW;QAGMC;IAFpC,MAAMA,IAAIrI,YAAYoI;IACtB,IAAI,CAACC,GAAG,MAAM,IAAIpI,WAAW,kBAAkB;IAC/C,MAAMqI,MAAM,GAAGD,EAAEE,IAAI,CAAC,CAAC,EAAEF,EAAEtE,KAAK,CAAC,CAAC,GAAEsE,SAAAA,EAAEjF,GAAG,cAALiF,oBAAAA,SAAS,IAAI;IACjD,IAAIG,IAAIP,UAAUQ,GAAG,CAACH;IACtB,IAAI,CAACE,GAAG;QACNA,IAAIH,EAAEE,IAAI,KAAK,QAAQvB,YAAYqB,EAAEtE,KAAK,EAAEsE,EAAEjF,GAAG,EAAEiF,EAAEjB,GAAG,IAAItD,eAAeuE,EAAEtE,KAAK;QAClFkE,UAAUS,GAAG,CAACJ,KAAKE;IACrB;IACA,OAAOA;AACT;AAEA,4EAA4E;AAC5E,SAASG,QAAQC,IAAkB,EAAEvD,IAAY,EAAEwD,IAAa;IAC9D,MAAMrC,IAAI,IAAIb,aAAaN;IAC3B,IAAIsB,OAAO;IACX,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKC,QAAQiC,IAAI,CAAClC,EAAE,GAAGkC,IAAI,CAAClC,EAAE;IACxDC,OAAOC,KAAKC,IAAI,CAACF,QAAQ;IACzB,IAAK,IAAID,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKF,CAAC,CAACE,EAAE,GAAGkC,IAAI,CAAClC,EAAE,GAAGC;IAChD,IAAI,CAACkC,MAAM,OAAO;QAAErC;QAAGsC,OAAO;IAAE;IAChC,IAAIC,MAAM;IACV,IAAK,IAAIrC,IAAI,GAAGA,IAAIrB,MAAMqB,IAAKqC,MAAMnC,KAAKmC,GAAG,CAACA,KAAKnC,KAAKoC,GAAG,CAACxC,CAAC,CAACE,EAAE;IAChE,OAAO;QAAEF;QAAGsC,OAAOC,MAAM,OAAO;IAAE;AACpC;AAEA,qFAAqF;AACrF,6CAA6C;AAC7C,OAAO,eAAeE,aAAa9G,EAAgB,EAAEiG,GAAW,EAAEc,OAAe;IAC/E,MAAMC,WAAW,MAAMhB,YAAYC,MAAM,kDAAkD;IAC3F,MAAMgB,QAAQjH,GAAGM,OAAO,CAAC,oFAAoF4G,GAAG;IAChH,IAAID,MAAMtI,MAAM,KAAK,GAAG;IACxB,MAAMwI,YAAY1C,KAAK2C,GAAG,CAACnJ,YAAY+I,SAAS9D,IAAI;IAEpD,MAAMmE,SAAS,IAAItB;IACnB,KAAK,MAAMuB,OAAOL,MAAO;YACVI;QAAb,MAAME,QAAOF,cAAAA,OAAOf,GAAG,CAACgB,IAAIjH,IAAI,eAAnBgH,yBAAAA,cAAwB,EAAE;QACvCE,KAAK1I,IAAI,CAACyI,IAAIE,KAAK;QACnBH,OAAOd,GAAG,CAACe,IAAIjH,IAAI,EAAEkH;IACvB;IAEA,MAAME,OAA6D,EAAE;IACrE,KAAK,MAAM,CAACpH,MAAMqH,UAAU,IAAIL,OAAQ;QACtC,IAAIjI;QACJ,IAAI;YACF,qFAAqF;YACrF,sFAAsF;YACtFA,SAASpB,UAAU;gBAAE2J,SAAStH;gBAAMuH,SAAShK,KAAKmJ,SAAS1G;gBAAOwH,SAAS;gBAAGC,MAAM;gBAAGC,SAAS,EAAE;gBAAElI,OAAO;YAAK,GAAG;gBAACA;aAAM,EAAEmI,GAAG,CAACvH,SAAS,CAACZ,KAAK;QACjJ,EAAE,OAAM;YACN,UAAU,gEAAgE;QAC5E;QACA,KAAK,MAAMe,OAAO8G,UAAW,IAAItI,MAAM,CAACwB,IAAI,EAAE6G,KAAK5I,IAAI,CAAC;YAAEwB;YAAMmH,OAAO5G;YAAKpB,MAAMJ,MAAM,CAACwB,IAAI,CAACpB,IAAI;QAAC;IACrG;IAEA,MAAMyI,SAASjI,GAAGM,OAAO,CAAC;IAC1B,wFAAwF;IACxF,6DAA6D;IAC7D,MAAM4H,SAASnK,SAAS,oBAAoB0J,KAAK9I,MAAM;IACvD,IAAK,IAAID,IAAI,GAAGA,IAAI+I,KAAK9I,MAAM,EAAED,KAAKR,MAAO;QAC3C,MAAMiK,QAAQV,KAAKhI,KAAK,CAACf,GAAGA,IAAIR;QAChC,MAAMkK,UAAU,MAAMpB,SAASnH,KAAK,CAACsI,MAAMvD,GAAG,CAAC,CAACyD,IAAMA,EAAE7I,IAAI;QAC5DQ,GAAGC,IAAI,CAAC;QACR,IAAI;YACFkI,MAAM9I,OAAO,CAAC,CAACiJ,KAAKD;gBAClB,MAAM,EAAEhE,CAAC,EAAEsC,KAAK,EAAE,GAAGH,QAAQ4B,OAAO,CAACC,EAAE,EAAElB,WAAW;gBACpD,MAAMoB,IAAI,IAAIC,UAAUrB;gBACxB,IAAK,IAAI5C,IAAI,GAAGA,IAAI4C,WAAW5C,IAAKgE,CAAC,CAAChE,EAAE,GAAGE,KAAKgE,KAAK,CAACpE,CAAC,CAACE,EAAE,GAAGoC;gBAC7DsB,OAAO1H,GAAG,CAACoG,OAAOnF,OAAOC,IAAI,CAAC8G,EAAE9E,MAAM,GAAG6E,IAAIjI,IAAI,EAAEiI,IAAId,KAAK;YAC9D;YACAxH,GAAGC,IAAI,CAAC;QACV,EAAE,OAAOyI,KAAK;YACZ1I,GAAGC,IAAI,CAAC;YACR,MAAMyI;QACR;QACAR,OAAOS,IAAI,CAAClE,KAAK2C,GAAG,CAAC1I,IAAIR,OAAOuJ,KAAK9I,MAAM;IAC7C;IACAuJ,OAAOU,MAAM;AACf;AAEA,oFAAoF;AACpF,wFAAwF;AACxF,oBAAoB;AACpB,4FAA4F;AAC5F,8FAA8F;AAC9F,2FAA2F;AAC3F,6FAA6F;AAC7F,8BAA8B;AAC9B,OAAO,eAAeC,mBAAmB7I,EAAgB,EAAEiG,GAAW,EAAE6C,KAAa,EAAE3H,MAAa;QAOpF2H;IANd,MAAM/B,UAAU,AAACd,IAAgCc,OAAO;IACxD,IAAI,CAACA,SAAS,MAAM,IAAIjJ,WAAW,eAAe;IAClD,MAAMgJ,aAAa9G,IAAIiG,KAAKc;IAE5B,MAAMC,WAAW,MAAMhB,YAAYC;IACnC,MAAMkB,YAAY1C,KAAK2C,GAAG,CAACnJ,YAAY+I,SAAS9D,IAAI;IACpD,MAAM1D,OAAO,EAACsJ,eAAAA,MAAMC,KAAK,CAAC,gCAAZD,0BAAAA,eAAkC,EAAE,EAAE5J,MAAM,CAAC,CAAC8J,IAAM,CAAC;YAAC;YAAO;YAAM;YAAO;SAAO,CAACC,QAAQ,CAACD,IAAIpL,IAAI,CAAC;IAClH,MAAM,EAAEyG,GAAG6E,EAAE,EAAE,GAAG1C,QAAQ,AAAC,CAAA,MAAMQ,SAASnH,KAAK,CAAC;QAACL;KAAK,CAAA,CAAE,CAAC,EAAE,EAAE2H,WAAW;IAExE,MAAMgC,OAAOnJ,GAAGM,OAAO,CAAC,+FAA+F4G,GAAG;IAQ1H,MAAMkC,OAAO,IAAIrD;IACjB,KAAK,MAAMuB,OAAO6B,KAAM;QACtB,MAAMZ,IAAI,IAAIC,UAAUlB,IAAI+B,MAAM,CAAC5F,MAAM,EAAE6D,IAAI+B,MAAM,CAAChG,UAAU,EAAEoB,KAAK2C,GAAG,CAACD,WAAWG,IAAI+B,MAAM,CAACC,UAAU;QAC3G,IAAIC,MAAM;QACV,IAAK,IAAIhF,IAAI,GAAGA,IAAIgE,EAAE5J,MAAM,EAAE4F,IAAKgF,OAAOhB,CAAC,CAAChE,EAAE,GAAG2E,EAAE,CAAC3E,EAAE;QACtD,MAAMiF,QAAQD,MAAMjC,IAAIX,KAAK;QAC7B,MAAM8C,WAAWL,KAAK9C,GAAG,CAACgB,IAAIjH,IAAI;QAClC,IAAI,CAACoJ,YAAYD,QAAQC,SAASD,KAAK,EAAEJ,KAAK7C,GAAG,CAACe,IAAIjH,IAAI,EAAE;YAAEmJ;YAAOlL,OAAO,CAAC,CAAC,EAAEgJ,IAAIoC,UAAU,CAAC,CAAC,EAAEpC,IAAIqC,QAAQ,EAAE;QAAC;IACnH;IACA,OAAO;WAAIP,KAAKvG,OAAO;KAAG,CACvB+G,IAAI,CAAC,CAACC,GAAGC,IAAMA,CAAC,CAAC,EAAE,CAACN,KAAK,GAAGK,CAAC,CAAC,EAAE,CAACL,KAAK,EACtC/J,KAAK,CAAC,GAAG0B,QACTyD,GAAG,CAAC,CAAC,CAACvE,MAAMyJ,EAAE,GAAM,CAAA;YAAEzJ;YAAM/B,OAAOwL,EAAExL,KAAK;YAAEyL,YAAYtF,KAAKgE,KAAK,CAACqB,EAAEN,KAAK,GAAG,QAAQ;QAAK,CAAA;AAC/F"}

package/dist/esm/features/links.js

@@ -1,5 +1,7 @@
 import posix from 'node:path/posix';
-// links(src, target, dst): target as written, dst the resolved path or NULL (a queryable dead link).
+// links(src, target, target_base, dst): target as written, target_base its baseKey
+// (indexed, for the incremental resolve below), dst the resolved path or NULL (a
+// queryable dead link).
 // Wikilinks ([[target]], [[target#anchor|alias]], embeds) plus relative markdown links to .md files.
 function extract(_raw, body) {
     const targets = new Set();
@@ -15,6 +17,9 @@ function extract(_raw, body) {
         ...targets
     ];
 }
+function cleanTarget(target) {
+    return target.replace(/\\/g, '/').replace(/^\.\//, '');
+}
 function baseKey(path) {
     return posix.basename(path).replace(/\.md$/i, '').toLowerCase();
 }
@@ -23,7 +28,7 @@ function baseKey(path) {
 function resolveTarget(src, target, pathSet, byBase) {
     var _ref;
     var _byBase_get;
-    const clean = target.replace(/\\/g, '/').replace(/^\.\//, '');
+    const clean = cleanTarget(target);
     const fromSrc = posix.normalize(posix.join(posix.dirname(src), clean));
     for (const candidate of [
         clean,
@@ -35,26 +40,74 @@ function resolveTarget(src, target, pathSet, byBase) {
     }
     return (_ref = (_byBase_get = byBase.get(baseKey(clean))) === null || _byBase_get === void 0 ? void 0 : _byBase_get[0]) !== null && _ref !== void 0 ? _ref : null;
 }
-// A new or deleted file can change any note's resolution, so re-resolve the whole table
-// whenever the tree changes. An in-memory probe per row; writes only on change.
-function resolveAll(db, files) {
-    const pathSet = new Set(files.map((f)=>f.relPath));
+function buildByBase(files) {
     const byBase = new Map();
-    for (const path of [
-        ...pathSet
-    ].sort()){
-        const list = byBase.get(baseKey(path));
+    for (const path of files.map((f)=>f.relPath).sort()){
+        const key = baseKey(path);
+        const list = byBase.get(key);
         if (list) list.push(path);
-        else byBase.set(baseKey(path), [
+        else byBase.set(key, [
             path
         ]);
     }
-    const rows = db.prepare('SELECT src, target, dst FROM links').all();
+    return byBase;
+}
+// Applies resolveTarget to `rows`, writing only rows whose dst actually changes.
+function resolveRows(db, rows, pathSet, byBase) {
     const update = db.prepare('UPDATE links SET dst = ? WHERE src = ? AND target = ?');
+    let changed = false;
     for (const row of rows){
         const dst = resolveTarget(row.src, row.target, pathSet, byBase);
-        if (dst !== row.dst) update.run(dst, row.src, row.target);
+        if (dst !== row.dst) {
+            update.run(dst, row.src, row.target);
+            changed = true;
+        }
     }
+    return changed;
+}
+// A new or deleted file can change any note's resolution, so re-resolve the whole table.
+// Fallback for cold builds and large deltas -- see afterReconcile's threshold.
+function resolveAll(db, files) {
+    const pathSet = new Set(files.map((f)=>f.relPath));
+    const byBase = buildByBase(files);
+    const rows = db.prepare('SELECT src, target, dst FROM links').all();
+    return resolveRows(db, rows, pathSet, byBase);
+}
+// Re-resolve only what this reconcile could have affected: (a) links whose src was just
+// re-stored (their rows are fresh, dst reset to NULL by store()), (b) links whose target's
+// basename matches a file that appeared or disappeared this reconcile (resolution and
+// lexicographic-tie ambiguity hinge on exactly that set), and (c) links whose dst was a
+// path that just vanished. All three are indexed lookups (src is the links PK's leading
+// column; target_base and dst each have their own index), so this reads only the rows this
+// reconcile could plausibly have changed, not the whole table.
+function resolveIncremental(db, delta) {
+    const pathSet = new Set(delta.files.map((f)=>f.relPath));
+    const byBase = buildByBase(delta.files);
+    const changedBasenames = new Set();
+    for (const p of delta.added)changedBasenames.add(baseKey(p));
+    for (const p of delta.vanished)changedBasenames.add(baseKey(p));
+    // \0 can't appear in a path or target, so the key never collides; a space can.
+    const candidates = new Map();
+    const collect = (rows)=>{
+        for (const row of rows)candidates.set(`${row.src}\0${row.target}`, row);
+    };
+    // Chunked so a large-but-under-threshold delta never exceeds SQLite's bound-variable
+    // limit (SQLITE_MAX_VARIABLE_NUMBER, 32766).
+    const collectIn = (column, keys)=>{
+        for(let i = 0; i < keys.length; i += 500){
+            const chunk = keys.slice(i, i + 500);
+            const placeholders = chunk.map(()=>'?').join(', ');
+            collect(db.prepare(`SELECT src, target, dst FROM links WHERE ${column} IN (${placeholders})`).all(...chunk));
+        }
+    };
+    collectIn('src', delta.reparsed);
+    collectIn('target_base', [
+        ...changedBasenames
+    ]);
+    collectIn('dst', delta.vanished);
+    return resolveRows(db, [
+        ...candidates.values()
+    ], pathSet, byBase);
 }
 // Resolved edges, for rank and for find's graph expansion.
 export function linkEdges(db) {
@@ -63,19 +116,76 @@ export function linkEdges(db) {
             r.dst
         ]);
 }
+// A removed path's outbound link rows are already gone by the time afterReconcile runs
+// (remove() below deletes them for vanished files AND for reparsed existing files), so
+// afterReconcile can't tell from the links table alone whether deleted rows carried edges.
+// remove() records that here, keyed on the reconcile's own delta object -- state dies with
+// the reconcile, including on ROLLBACK, and distinct reconciles never share it.
+const removedWithLinks = new WeakMap();
+function recordRemoved(delta, path) {
+    let set = removedWithLinks.get(delta);
+    if (!set) {
+        set = new Set();
+        removedWithLinks.set(delta, set);
+    }
+    set.add(path);
+}
+// remove() runs per path; the vanished list is an array, so cache the Set per delta.
+const vanishedSets = new WeakMap();
+function vanishedSet(delta) {
+    let set = vanishedSets.get(delta);
+    if (!set) {
+        set = new Set(delta.vanished);
+        vanishedSets.set(delta, set);
+    }
+    return set;
+}
 export const links = {
     name: 'links',
     schema (db) {
-        db.exec('CREATE TABLE IF NOT EXISTS links (src TEXT, target TEXT, dst TEXT, PRIMARY KEY (src, target))');
+        db.exec('CREATE TABLE IF NOT EXISTS links (src TEXT, target TEXT, target_base TEXT, dst TEXT, PRIMARY KEY (src, target))');
         db.exec('CREATE INDEX IF NOT EXISTS links_dst ON links(dst)');
+        db.exec('CREATE INDEX IF NOT EXISTS links_target_base ON links(target_base)');
     },
     extract,
-    remove (db, path) {
-        db.prepare('DELETE FROM links WHERE src = ?').run(path);
+    // Reparsed docs are diffed in store() below instead of wiped here: deleting and
+    // re-inserting resets dst to NULL, which made every touch-only reparse read as an edge
+    // change and recompute PageRank -- measured as most of the remaining update cost.
+    remove (db, path, delta) {
+        if (!vanishedSet(delta).has(path)) return;
+        const result = db.prepare('DELETE FROM links WHERE src = ?').run(path);
+        if (Number(result.changes) > 0) recordRemoved(delta, path);
     },
-    store (db, path, extracted) {
-        const insert = db.prepare('INSERT OR REPLACE INTO links (src, target, dst) VALUES (?, ?, NULL)');
-        for (const target of extracted)insert.run(path, target);
+    store (db, path, extracted, delta) {
+        const targets = extracted;
+        // Stale rows (targets the new parse no longer contains) are real edge removals.
+        let stale;
+        if (targets.length === 0) {
+            stale = db.prepare('DELETE FROM links WHERE src = ?').run(path);
+        } else {
+            const placeholders = targets.map(()=>'?').join(', ');
+            stale = db.prepare(`DELETE FROM links WHERE src = ? AND target NOT IN (${placeholders})`).run(path, ...targets);
+        }
+        if (Number(stale.changes) > 0) recordRemoved(delta, path);
+        // Upsert preserves dst on surviving rows, so an unchanged link resolves to the same
+        // value and reports no change; only genuinely new rows start at NULL.
+        const insert = db.prepare('INSERT INTO links (src, target, target_base, dst) VALUES (?, ?, ?, NULL) ON CONFLICT(src, target) DO UPDATE SET target_base = excluded.target_base');
+        for (const target of targets)insert.run(path, target, baseKey(cleanTarget(target)));
     },
-    afterReconcile: resolveAll
+    afterReconcile (db, delta) {
+        var _ref;
+        var _removedWithLinks_get;
+        // Any deleted rows that carried edges mean the edge set shrank -- whether the file
+        // vanished or was reparsed down to fewer/no links. dstChanged alone misses the
+        // reparsed-to-zero-links case (no surviving rows to re-resolve).
+        const removeHadLinks = ((_ref = (_removedWithLinks_get = removedWithLinks.get(delta)) === null || _removedWithLinks_get === void 0 ? void 0 : _removedWithLinks_get.size) !== null && _ref !== void 0 ? _ref : 0) > 0;
+        const churn = delta.reparsed.length + delta.vanished.length;
+        // Fallback rule: correctness over cleverness once too much of the tree moved in one
+        // reconcile -- a full pass is the one guaranteed to match resolveTarget's ambiguity
+        // rules exactly. A cold build reparses ~every file, so it clears this threshold on its
+        // own without a separate check.
+        const large = delta.files.length === 0 || churn > 0.2 * delta.files.length;
+        const dstChanged = large ? resolveAll(db, delta.files) : resolveIncremental(db, delta);
+        delta.linksChanged = dstChanged || removeHadLinks;
+    }
 };

package/dist/esm/features/links.js.map

@@ -1 +1 @@
-{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/links.ts"],"sourcesContent":["import posix from 'node:path/posix';\nimport type { DatabaseSync } from 'node:sqlite';\nimport type { FileStat } from '../scan.ts';\nimport type { Feature } from './types.ts';\n\n// links(src, target, dst): target as written, dst the resolved path or NULL (a queryable dead link).\n\n// Wikilinks ([[target]], [[target#anchor|alias]], embeds) plus relative markdown links to .md files.\nfunction extract(_raw: string, body: string): string[] {\n  const targets = new Set<string>();\n  for (const m of body.matchAll(/\\[\\[([^\\]|#]+)(?:#[^\\]|]*)?(?:\\|[^\\]]*)?\\]\\]/g)) {\n    const target = m[1].trim();\n    if (target) targets.add(target);\n  }\n  for (const m of body.matchAll(/\\]\\(([^)]+\\.md)(?:#[^)]*)?\\)/g)) {\n    const target = m[1].trim();\n    if (target && !/^[a-z]+:\\/\\//i.test(target)) targets.add(target);\n  }\n  return [...targets];\n}\n\nfunction baseKey(path: string): string {\n  return posix.basename(path).replace(/\\.md$/i, '').toLowerCase();\n}\n\n// Obsidian-style: exact relative path (with/without .md), path relative to the linking\n// note's directory, then basename match (lexicographically first on ties).\nfunction resolveTarget(src: string, target: string, pathSet: Set<string>, byBase: Map<string, string[]>): string | null {\n  const clean = target.replace(/\\\\/g, '/').replace(/^\\.\\//, '');\n  const fromSrc = posix.normalize(posix.join(posix.dirname(src), clean));\n  for (const candidate of [clean, `${clean}.md`, fromSrc, `${fromSrc}.md`]) {\n    if (pathSet.has(candidate)) return candidate;\n  }\n  return byBase.get(baseKey(clean))?.[0] ?? null;\n}\n\n// A new or deleted file can change any note's resolution, so re-resolve the whole table\n// whenever the tree changes. An in-memory probe per row; writes only on change.\nfunction resolveAll(db: DatabaseSync, files: FileStat[]): void {\n  const pathSet = new Set(files.map((f) => f.relPath));\n  const byBase = new Map<string, string[]>();\n  for (const path of [...pathSet].sort()) {\n    const list = byBase.get(baseKey(path));\n    if (list) list.push(path);\n    else byBase.set(baseKey(path), [path]);\n  }\n  const rows = db.prepare('SELECT src, target, dst FROM links').all() as Array<{ src: string; target: string; dst: string | null }>;\n  const update = db.prepare('UPDATE links SET dst = ? WHERE src = ? AND target = ?');\n  for (const row of rows) {\n    const dst = resolveTarget(row.src, row.target, pathSet, byBase);\n    if (dst !== row.dst) update.run(dst, row.src, row.target);\n  }\n}\n\n// Resolved edges, for rank and for find's graph expansion.\nexport function linkEdges(db: DatabaseSync): [string, string][] {\n  return (db.prepare('SELECT src, dst FROM links WHERE dst IS NOT NULL').all() as Array<{ src: string; dst: string }>).map((r) => [r.src, r.dst]);\n}\n\nexport const links: Feature = {\n  name: 'links',\n  schema(db) {\n    db.exec('CREATE TABLE IF NOT EXISTS links (src TEXT, target TEXT, dst TEXT, PRIMARY KEY (src, target))');\n    db.exec('CREATE INDEX IF NOT EXISTS links_dst ON links(dst)');\n  },\n  extract,\n  remove(db, path) {\n    db.prepare('DELETE FROM links WHERE src = ?').run(path);\n  },\n  store(db, path, extracted) {\n    const insert = db.prepare('INSERT OR REPLACE INTO links (src, target, dst) VALUES (?, ?, NULL)');\n    for (const target of extracted as string[]) insert.run(path, target);\n  },\n  afterReconcile: resolveAll,\n};\n"],"names":["posix","extract","_raw","body","targets","Set","m","matchAll","target","trim","add","test","baseKey","path","basename","replace","toLowerCase","resolveTarget","src","pathSet","byBase","clean","fromSrc","normalize","join","dirname","candidate","has","get","resolveAll","db","files","map","f","relPath","Map","sort","list","push","set","rows","prepare","all","update","row","dst","run","linkEdges","r","links","name","schema","exec","remove","store","extracted","insert","afterReconcile"],"mappings":"AAAA,OAAOA,WAAW,kBAAkB;AAKpC,qGAAqG;AAErG,qGAAqG;AACrG,SAASC,QAAQC,IAAY,EAAEC,IAAY;IACzC,MAAMC,UAAU,IAAIC;IACpB,KAAK,MAAMC,KAAKH,KAAKI,QAAQ,CAAC,iDAAkD;QAC9E,MAAMC,SAASF,CAAC,CAAC,EAAE,CAACG,IAAI;QACxB,IAAID,QAAQJ,QAAQM,GAAG,CAACF;IAC1B;IACA,KAAK,MAAMF,KAAKH,KAAKI,QAAQ,CAAC,iCAAkC;QAC9D,MAAMC,SAASF,CAAC,CAAC,EAAE,CAACG,IAAI;QACxB,IAAID,UAAU,CAAC,gBAAgBG,IAAI,CAACH,SAASJ,QAAQM,GAAG,CAACF;IAC3D;IACA,OAAO;WAAIJ;KAAQ;AACrB;AAEA,SAASQ,QAAQC,IAAY;IAC3B,OAAOb,MAAMc,QAAQ,CAACD,MAAME,OAAO,CAAC,UAAU,IAAIC,WAAW;AAC/D;AAEA,uFAAuF;AACvF,2EAA2E;AAC3E,SAASC,cAAcC,GAAW,EAAEV,MAAc,EAAEW,OAAoB,EAAEC,MAA6B;;QAM9FA;IALP,MAAMC,QAAQb,OAAOO,OAAO,CAAC,OAAO,KAAKA,OAAO,CAAC,SAAS;IAC1D,MAAMO,UAAUtB,MAAMuB,SAAS,CAACvB,MAAMwB,IAAI,CAACxB,MAAMyB,OAAO,CAACP,MAAMG;IAC/D,KAAK,MAAMK,aAAa;QAACL;QAAO,GAAGA,MAAM,GAAG,CAAC;QAAEC;QAAS,GAAGA,QAAQ,GAAG,CAAC;KAAC,CAAE;QACxE,IAAIH,QAAQQ,GAAG,CAACD,YAAY,OAAOA;IACrC;IACA,gBAAON,cAAAA,OAAOQ,GAAG,CAAChB,QAAQS,qBAAnBD,kCAAAA,WAA4B,CAAC,EAAE,uCAAI;AAC5C;AAEA,wFAAwF;AACxF,gFAAgF;AAChF,SAASS,WAAWC,EAAgB,EAAEC,KAAiB;IACrD,MAAMZ,UAAU,IAAId,IAAI0B,MAAMC,GAAG,CAAC,CAACC,IAAMA,EAAEC,OAAO;IAClD,MAAMd,SAAS,IAAIe;IACnB,KAAK,MAAMtB,QAAQ;WAAIM;KAAQ,CAACiB,IAAI,GAAI;QACtC,MAAMC,OAAOjB,OAAOQ,GAAG,CAAChB,QAAQC;QAChC,IAAIwB,MAAMA,KAAKC,IAAI,CAACzB;aACfO,OAAOmB,GAAG,CAAC3B,QAAQC,OAAO;YAACA;SAAK;IACvC;IACA,MAAM2B,OAAOV,GAAGW,OAAO,CAAC,sCAAsCC,GAAG;IACjE,MAAMC,SAASb,GAAGW,OAAO,CAAC;IAC1B,KAAK,MAAMG,OAAOJ,KAAM;QACtB,MAAMK,MAAM5B,cAAc2B,IAAI1B,GAAG,EAAE0B,IAAIpC,MAAM,EAAEW,SAASC;QACxD,IAAIyB,QAAQD,IAAIC,GAAG,EAAEF,OAAOG,GAAG,CAACD,KAAKD,IAAI1B,GAAG,EAAE0B,IAAIpC,MAAM;IAC1D;AACF;AAEA,2DAA2D;AAC3D,OAAO,SAASuC,UAAUjB,EAAgB;IACxC,OAAO,AAACA,GAAGW,OAAO,CAAC,oDAAoDC,GAAG,GAA2CV,GAAG,CAAC,CAACgB,IAAM;YAACA,EAAE9B,GAAG;YAAE8B,EAAEH,GAAG;SAAC;AAChJ;AAEA,OAAO,MAAMI,QAAiB;IAC5BC,MAAM;IACNC,QAAOrB,EAAE;QACPA,GAAGsB,IAAI,CAAC;QACRtB,GAAGsB,IAAI,CAAC;IACV;IACAnD;IACAoD,QAAOvB,EAAE,EAAEjB,IAAI;QACbiB,GAAGW,OAAO,CAAC,mCAAmCK,GAAG,CAACjC;IACpD;IACAyC,OAAMxB,EAAE,EAAEjB,IAAI,EAAE0C,SAAS;QACvB,MAAMC,SAAS1B,GAAGW,OAAO,CAAC;QAC1B,KAAK,MAAMjC,UAAU+C,UAAuBC,OAAOV,GAAG,CAACjC,MAAML;IAC/D;IACAiD,gBAAgB5B;AAClB,EAAE"}
+{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/links.ts"],"sourcesContent":["import posix from 'node:path/posix';\nimport type { DatabaseSync } from 'node:sqlite';\nimport type { FileStat } from '../scan.ts';\nimport type { Feature, ReconcileDelta } from './types.ts';\n\n// links(src, target, target_base, dst): target as written, target_base its baseKey\n// (indexed, for the incremental resolve below), dst the resolved path or NULL (a\n// queryable dead link).\n\n// Wikilinks ([[target]], [[target#anchor|alias]], embeds) plus relative markdown links to .md files.\nfunction extract(_raw: string, body: string): string[] {\n  const targets = new Set<string>();\n  for (const m of body.matchAll(/\\[\\[([^\\]|#]+)(?:#[^\\]|]*)?(?:\\|[^\\]]*)?\\]\\]/g)) {\n    const target = m[1].trim();\n    if (target) targets.add(target);\n  }\n  for (const m of body.matchAll(/\\]\\(([^)]+\\.md)(?:#[^)]*)?\\)/g)) {\n    const target = m[1].trim();\n    if (target && !/^[a-z]+:\\/\\//i.test(target)) targets.add(target);\n  }\n  return [...targets];\n}\n\nfunction cleanTarget(target: string): string {\n  return target.replace(/\\\\/g, '/').replace(/^\\.\\//, '');\n}\n\nfunction baseKey(path: string): string {\n  return posix.basename(path).replace(/\\.md$/i, '').toLowerCase();\n}\n\n// Obsidian-style: exact relative path (with/without .md), path relative to the linking\n// note's directory, then basename match (lexicographically first on ties).\nfunction resolveTarget(src: string, target: string, pathSet: Set<string>, byBase: Map<string, string[]>): string | null {\n  const clean = cleanTarget(target);\n  const fromSrc = posix.normalize(posix.join(posix.dirname(src), clean));\n  for (const candidate of [clean, `${clean}.md`, fromSrc, `${fromSrc}.md`]) {\n    if (pathSet.has(candidate)) return candidate;\n  }\n  return byBase.get(baseKey(clean))?.[0] ?? null;\n}\n\nfunction buildByBase(files: FileStat[]): Map<string, string[]> {\n  const byBase = new Map<string, string[]>();\n  for (const path of files.map((f) => f.relPath).sort()) {\n    const key = baseKey(path);\n    const list = byBase.get(key);\n    if (list) list.push(path);\n    else byBase.set(key, [path]);\n  }\n  return byBase;\n}\n\ninterface LinkRow {\n  src: string;\n  target: string;\n  dst: string | null;\n}\n\n// Applies resolveTarget to `rows`, writing only rows whose dst actually changes.\nfunction resolveRows(db: DatabaseSync, rows: LinkRow[], pathSet: Set<string>, byBase: Map<string, string[]>): boolean {\n  const update = db.prepare('UPDATE links SET dst = ? WHERE src = ? AND target = ?');\n  let changed = false;\n  for (const row of rows) {\n    const dst = resolveTarget(row.src, row.target, pathSet, byBase);\n    if (dst !== row.dst) {\n      update.run(dst, row.src, row.target);\n      changed = true;\n    }\n  }\n  return changed;\n}\n\n// A new or deleted file can change any note's resolution, so re-resolve the whole table.\n// Fallback for cold builds and large deltas -- see afterReconcile's threshold.\nfunction resolveAll(db: DatabaseSync, files: FileStat[]): boolean {\n  const pathSet = new Set(files.map((f) => f.relPath));\n  const byBase = buildByBase(files);\n  const rows = db.prepare('SELECT src, target, dst FROM links').all() as unknown as LinkRow[];\n  return resolveRows(db, rows, pathSet, byBase);\n}\n\n// Re-resolve only what this reconcile could have affected: (a) links whose src was just\n// re-stored (their rows are fresh, dst reset to NULL by store()), (b) links whose target's\n// basename matches a file that appeared or disappeared this reconcile (resolution and\n// lexicographic-tie ambiguity hinge on exactly that set), and (c) links whose dst was a\n// path that just vanished. All three are indexed lookups (src is the links PK's leading\n// column; target_base and dst each have their own index), so this reads only the rows this\n// reconcile could plausibly have changed, not the whole table.\nfunction resolveIncremental(db: DatabaseSync, delta: ReconcileDelta): boolean {\n  const pathSet = new Set(delta.files.map((f) => f.relPath));\n  const byBase = buildByBase(delta.files);\n\n  const changedBasenames = new Set<string>();\n  for (const p of delta.added) changedBasenames.add(baseKey(p));\n  for (const p of delta.vanished) changedBasenames.add(baseKey(p));\n\n  // \\0 can't appear in a path or target, so the key never collides; a space can.\n  const candidates = new Map<string, LinkRow>();\n  const collect = (rows: LinkRow[]) => {\n    for (const row of rows) candidates.set(`${row.src}\\0${row.target}`, row);\n  };\n  // Chunked so a large-but-under-threshold delta never exceeds SQLite's bound-variable\n  // limit (SQLITE_MAX_VARIABLE_NUMBER, 32766).\n  const collectIn = (column: string, keys: string[]) => {\n    for (let i = 0; i < keys.length; i += 500) {\n      const chunk = keys.slice(i, i + 500);\n      const placeholders = chunk.map(() => '?').join(', ');\n      collect(db.prepare(`SELECT src, target, dst FROM links WHERE ${column} IN (${placeholders})`).all(...chunk) as unknown as LinkRow[]);\n    }\n  };\n\n  collectIn('src', delta.reparsed);\n  collectIn('target_base', [...changedBasenames]);\n  collectIn('dst', delta.vanished);\n\n  return resolveRows(db, [...candidates.values()], pathSet, byBase);\n}\n\n// Resolved edges, for rank and for find's graph expansion.\nexport function linkEdges(db: DatabaseSync): [string, string][] {\n  return (db.prepare('SELECT src, dst FROM links WHERE dst IS NOT NULL').all() as Array<{ src: string; dst: string }>).map((r) => [r.src, r.dst]);\n}\n\n// A removed path's outbound link rows are already gone by the time afterReconcile runs\n// (remove() below deletes them for vanished files AND for reparsed existing files), so\n// afterReconcile can't tell from the links table alone whether deleted rows carried edges.\n// remove() records that here, keyed on the reconcile's own delta object -- state dies with\n// the reconcile, including on ROLLBACK, and distinct reconciles never share it.\nconst removedWithLinks = new WeakMap<ReconcileDelta, Set<string>>();\n\nfunction recordRemoved(delta: ReconcileDelta, path: string): void {\n  let set = removedWithLinks.get(delta);\n  if (!set) {\n    set = new Set();\n    removedWithLinks.set(delta, set);\n  }\n  set.add(path);\n}\n\n// remove() runs per path; the vanished list is an array, so cache the Set per delta.\nconst vanishedSets = new WeakMap<ReconcileDelta, Set<string>>();\nfunction vanishedSet(delta: ReconcileDelta): Set<string> {\n  let set = vanishedSets.get(delta);\n  if (!set) {\n    set = new Set(delta.vanished);\n    vanishedSets.set(delta, set);\n  }\n  return set;\n}\n\nexport const links: Feature = {\n  name: 'links',\n  schema(db) {\n    db.exec('CREATE TABLE IF NOT EXISTS links (src TEXT, target TEXT, target_base TEXT, dst TEXT, PRIMARY KEY (src, target))');\n    db.exec('CREATE INDEX IF NOT EXISTS links_dst ON links(dst)');\n    db.exec('CREATE INDEX IF NOT EXISTS links_target_base ON links(target_base)');\n  },\n  extract,\n  // Reparsed docs are diffed in store() below instead of wiped here: deleting and\n  // re-inserting resets dst to NULL, which made every touch-only reparse read as an edge\n  // change and recompute PageRank -- measured as most of the remaining update cost.\n  remove(db, path, delta) {\n    if (!vanishedSet(delta).has(path)) return;\n    const result = db.prepare('DELETE FROM links WHERE src = ?').run(path);\n    if (Number(result.changes) > 0) recordRemoved(delta, path);\n  },\n  store(db, path, extracted, delta) {\n    const targets = extracted as string[];\n    // Stale rows (targets the new parse no longer contains) are real edge removals.\n    let stale: ReturnType<ReturnType<DatabaseSync['prepare']>['run']>;\n    if (targets.length === 0) {\n      stale = db.prepare('DELETE FROM links WHERE src = ?').run(path);\n    } else {\n      const placeholders = targets.map(() => '?').join(', ');\n      stale = db.prepare(`DELETE FROM links WHERE src = ? AND target NOT IN (${placeholders})`).run(path, ...targets);\n    }\n    if (Number(stale.changes) > 0) recordRemoved(delta, path);\n    // Upsert preserves dst on surviving rows, so an unchanged link resolves to the same\n    // value and reports no change; only genuinely new rows start at NULL.\n    const insert = db.prepare('INSERT INTO links (src, target, target_base, dst) VALUES (?, ?, ?, NULL) ON CONFLICT(src, target) DO UPDATE SET target_base = excluded.target_base');\n    for (const target of targets) insert.run(path, target, baseKey(cleanTarget(target)));\n  },\n  afterReconcile(db, delta) {\n    // Any deleted rows that carried edges mean the edge set shrank -- whether the file\n    // vanished or was reparsed down to fewer/no links. dstChanged alone misses the\n    // reparsed-to-zero-links case (no surviving rows to re-resolve).\n    const removeHadLinks = (removedWithLinks.get(delta)?.size ?? 0) > 0;\n\n    const churn = delta.reparsed.length + delta.vanished.length;\n    // Fallback rule: correctness over cleverness once too much of the tree moved in one\n    // reconcile -- a full pass is the one guaranteed to match resolveTarget's ambiguity\n    // rules exactly. A cold build reparses ~every file, so it clears this threshold on its\n    // own without a separate check.\n    const large = delta.files.length === 0 || churn > 0.2 * delta.files.length;\n\n    const dstChanged = large ? resolveAll(db, delta.files) : resolveIncremental(db, delta);\n    delta.linksChanged = dstChanged || removeHadLinks;\n  },\n};\n"],"names":["posix","extract","_raw","body","targets","Set","m","matchAll","target","trim","add","test","cleanTarget","replace","baseKey","path","basename","toLowerCase","resolveTarget","src","pathSet","byBase","clean","fromSrc","normalize","join","dirname","candidate","has","get","buildByBase","files","Map","map","f","relPath","sort","key","list","push","set","resolveRows","db","rows","update","prepare","changed","row","dst","run","resolveAll","all","resolveIncremental","delta","changedBasenames","p","added","vanished","candidates","collect","collectIn","column","keys","i","length","chunk","slice","placeholders","reparsed","values","linkEdges","r","removedWithLinks","WeakMap","recordRemoved","vanishedSets","vanishedSet","links","name","schema","exec","remove","result","Number","changes","store","extracted","stale","insert","afterReconcile","removeHadLinks","size","churn","large","dstChanged","linksChanged"],"mappings":"AAAA,OAAOA,WAAW,kBAAkB;AAKpC,mFAAmF;AACnF,iFAAiF;AACjF,wBAAwB;AAExB,qGAAqG;AACrG,SAASC,QAAQC,IAAY,EAAEC,IAAY;IACzC,MAAMC,UAAU,IAAIC;IACpB,KAAK,MAAMC,KAAKH,KAAKI,QAAQ,CAAC,iDAAkD;QAC9E,MAAMC,SAASF,CAAC,CAAC,EAAE,CAACG,IAAI;QACxB,IAAID,QAAQJ,QAAQM,GAAG,CAACF;IAC1B;IACA,KAAK,MAAMF,KAAKH,KAAKI,QAAQ,CAAC,iCAAkC;QAC9D,MAAMC,SAASF,CAAC,CAAC,EAAE,CAACG,IAAI;QACxB,IAAID,UAAU,CAAC,gBAAgBG,IAAI,CAACH,SAASJ,QAAQM,GAAG,CAACF;IAC3D;IACA,OAAO;WAAIJ;KAAQ;AACrB;AAEA,SAASQ,YAAYJ,MAAc;IACjC,OAAOA,OAAOK,OAAO,CAAC,OAAO,KAAKA,OAAO,CAAC,SAAS;AACrD;AAEA,SAASC,QAAQC,IAAY;IAC3B,OAAOf,MAAMgB,QAAQ,CAACD,MAAMF,OAAO,CAAC,UAAU,IAAII,WAAW;AAC/D;AAEA,uFAAuF;AACvF,2EAA2E;AAC3E,SAASC,cAAcC,GAAW,EAAEX,MAAc,EAAEY,OAAoB,EAAEC,MAA6B;;QAM9FA;IALP,MAAMC,QAAQV,YAAYJ;IAC1B,MAAMe,UAAUvB,MAAMwB,SAAS,CAACxB,MAAMyB,IAAI,CAACzB,MAAM0B,OAAO,CAACP,MAAMG;IAC/D,KAAK,MAAMK,aAAa;QAACL;QAAO,GAAGA,MAAM,GAAG,CAAC;QAAEC;QAAS,GAAGA,QAAQ,GAAG,CAAC;KAAC,CAAE;QACxE,IAAIH,QAAQQ,GAAG,CAACD,YAAY,OAAOA;IACrC;IACA,gBAAON,cAAAA,OAAOQ,GAAG,CAACf,QAAQQ,qBAAnBD,kCAAAA,WAA4B,CAAC,EAAE,uCAAI;AAC5C;AAEA,SAASS,YAAYC,KAAiB;IACpC,MAAMV,SAAS,IAAIW;IACnB,KAAK,MAAMjB,QAAQgB,MAAME,GAAG,CAAC,CAACC,IAAMA,EAAEC,OAAO,EAAEC,IAAI,GAAI;QACrD,MAAMC,MAAMvB,QAAQC;QACpB,MAAMuB,OAAOjB,OAAOQ,GAAG,CAACQ;QACxB,IAAIC,MAAMA,KAAKC,IAAI,CAACxB;aACfM,OAAOmB,GAAG,CAACH,KAAK;YAACtB;SAAK;IAC7B;IACA,OAAOM;AACT;AAQA,iFAAiF;AACjF,SAASoB,YAAYC,EAAgB,EAAEC,IAAe,EAAEvB,OAAoB,EAAEC,MAA6B;IACzG,MAAMuB,SAASF,GAAGG,OAAO,CAAC;IAC1B,IAAIC,UAAU;IACd,KAAK,MAAMC,OAAOJ,KAAM;QACtB,MAAMK,MAAM9B,cAAc6B,IAAI5B,GAAG,EAAE4B,IAAIvC,MAAM,EAAEY,SAASC;QACxD,IAAI2B,QAAQD,IAAIC,GAAG,EAAE;YACnBJ,OAAOK,GAAG,CAACD,KAAKD,IAAI5B,GAAG,EAAE4B,IAAIvC,MAAM;YACnCsC,UAAU;QACZ;IACF;IACA,OAAOA;AACT;AAEA,yFAAyF;AACzF,+EAA+E;AAC/E,SAASI,WAAWR,EAAgB,EAAEX,KAAiB;IACrD,MAAMX,UAAU,IAAIf,IAAI0B,MAAME,GAAG,CAAC,CAACC,IAAMA,EAAEC,OAAO;IAClD,MAAMd,SAASS,YAAYC;IAC3B,MAAMY,OAAOD,GAAGG,OAAO,CAAC,sCAAsCM,GAAG;IACjE,OAAOV,YAAYC,IAAIC,MAAMvB,SAASC;AACxC;AAEA,wFAAwF;AACxF,2FAA2F;AAC3F,sFAAsF;AACtF,wFAAwF;AACxF,wFAAwF;AACxF,2FAA2F;AAC3F,+DAA+D;AAC/D,SAAS+B,mBAAmBV,EAAgB,EAAEW,KAAqB;IACjE,MAAMjC,UAAU,IAAIf,IAAIgD,MAAMtB,KAAK,CAACE,GAAG,CAAC,CAACC,IAAMA,EAAEC,OAAO;IACxD,MAAMd,SAASS,YAAYuB,MAAMtB,KAAK;IAEtC,MAAMuB,mBAAmB,IAAIjD;IAC7B,KAAK,MAAMkD,KAAKF,MAAMG,KAAK,CAAEF,iBAAiB5C,GAAG,CAACI,QAAQyC;IAC1D,KAAK,MAAMA,KAAKF,MAAMI,QAAQ,CAAEH,iBAAiB5C,GAAG,CAACI,QAAQyC;IAE7D,+EAA+E;IAC/E,MAAMG,aAAa,IAAI1B;IACvB,MAAM2B,UAAU,CAAChB;QACf,KAAK,MAAMI,OAAOJ,KAAMe,WAAWlB,GAAG,CAAC,GAAGO,IAAI5B,GAAG,CAAC,EAAE,EAAE4B,IAAIvC,MAAM,EAAE,EAAEuC;IACtE;IACA,qFAAqF;IACrF,6CAA6C;IAC7C,MAAMa,YAAY,CAACC,QAAgBC;QACjC,IAAK,IAAIC,IAAI,GAAGA,IAAID,KAAKE,MAAM,EAAED,KAAK,IAAK;YACzC,MAAME,QAAQH,KAAKI,KAAK,CAACH,GAAGA,IAAI;YAChC,MAAMI,eAAeF,MAAMhC,GAAG,CAAC,IAAM,KAAKR,IAAI,CAAC;YAC/CkC,QAAQjB,GAAGG,OAAO,CAAC,CAAC,yCAAyC,EAAEgB,OAAO,KAAK,EAAEM,aAAa,CAAC,CAAC,EAAEhB,GAAG,IAAIc;QACvG;IACF;IAEAL,UAAU,OAAOP,MAAMe,QAAQ;IAC/BR,UAAU,eAAe;WAAIN;KAAiB;IAC9CM,UAAU,OAAOP,MAAMI,QAAQ;IAE/B,OAAOhB,YAAYC,IAAI;WAAIgB,WAAWW,MAAM;KAAG,EAAEjD,SAASC;AAC5D;AAEA,2DAA2D;AAC3D,OAAO,SAASiD,UAAU5B,EAAgB;IACxC,OAAO,AAACA,GAAGG,OAAO,CAAC,oDAAoDM,GAAG,GAA2ClB,GAAG,CAAC,CAACsC,IAAM;YAACA,EAAEpD,GAAG;YAAEoD,EAAEvB,GAAG;SAAC;AAChJ;AAEA,uFAAuF;AACvF,uFAAuF;AACvF,2FAA2F;AAC3F,2FAA2F;AAC3F,gFAAgF;AAChF,MAAMwB,mBAAmB,IAAIC;AAE7B,SAASC,cAAcrB,KAAqB,EAAEtC,IAAY;IACxD,IAAIyB,MAAMgC,iBAAiB3C,GAAG,CAACwB;IAC/B,IAAI,CAACb,KAAK;QACRA,MAAM,IAAInC;QACVmE,iBAAiBhC,GAAG,CAACa,OAAOb;IAC9B;IACAA,IAAI9B,GAAG,CAACK;AACV;AAEA,qFAAqF;AACrF,MAAM4D,eAAe,IAAIF;AACzB,SAASG,YAAYvB,KAAqB;IACxC,IAAIb,MAAMmC,aAAa9C,GAAG,CAACwB;IAC3B,IAAI,CAACb,KAAK;QACRA,MAAM,IAAInC,IAAIgD,MAAMI,QAAQ;QAC5BkB,aAAanC,GAAG,CAACa,OAAOb;IAC1B;IACA,OAAOA;AACT;AAEA,OAAO,MAAMqC,QAAiB;IAC5BC,MAAM;IACNC,QAAOrC,EAAE;QACPA,GAAGsC,IAAI,CAAC;QACRtC,GAAGsC,IAAI,CAAC;QACRtC,GAAGsC,IAAI,CAAC;IACV;IACA/E;IACA,gFAAgF;IAChF,uFAAuF;IACvF,kFAAkF;IAClFgF,QAAOvC,EAAE,EAAE3B,IAAI,EAAEsC,KAAK;QACpB,IAAI,CAACuB,YAAYvB,OAAOzB,GAAG,CAACb,OAAO;QACnC,MAAMmE,SAASxC,GAAGG,OAAO,CAAC,mCAAmCI,GAAG,CAAClC;QACjE,IAAIoE,OAAOD,OAAOE,OAAO,IAAI,GAAGV,cAAcrB,OAAOtC;IACvD;IACAsE,OAAM3C,EAAE,EAAE3B,IAAI,EAAEuE,SAAS,EAAEjC,KAAK;QAC9B,MAAMjD,UAAUkF;QAChB,gFAAgF;QAChF,IAAIC;QACJ,IAAInF,QAAQ4D,MAAM,KAAK,GAAG;YACxBuB,QAAQ7C,GAAGG,OAAO,CAAC,mCAAmCI,GAAG,CAAClC;QAC5D,OAAO;YACL,MAAMoD,eAAe/D,QAAQ6B,GAAG,CAAC,IAAM,KAAKR,IAAI,CAAC;YACjD8D,QAAQ7C,GAAGG,OAAO,CAAC,CAAC,mDAAmD,EAAEsB,aAAa,CAAC,CAAC,EAAElB,GAAG,CAAClC,SAASX;QACzG;QACA,IAAI+E,OAAOI,MAAMH,OAAO,IAAI,GAAGV,cAAcrB,OAAOtC;QACpD,oFAAoF;QACpF,sEAAsE;QACtE,MAAMyE,SAAS9C,GAAGG,OAAO,CAAC;QAC1B,KAAK,MAAMrC,UAAUJ,QAASoF,OAAOvC,GAAG,CAAClC,MAAMP,QAAQM,QAAQF,YAAYJ;IAC7E;IACAiF,gBAAe/C,EAAE,EAAEW,KAAK;;YAIEmB;QAHxB,mFAAmF;QACnF,+EAA+E;QAC/E,iEAAiE;QACjE,MAAMkB,iBAAiB,UAAClB,wBAAAA,iBAAiB3C,GAAG,CAACwB,oBAArBmB,4CAAAA,sBAA6BmB,IAAI,uCAAI,KAAK;QAElE,MAAMC,QAAQvC,MAAMe,QAAQ,CAACJ,MAAM,GAAGX,MAAMI,QAAQ,CAACO,MAAM;QAC3D,oFAAoF;QACpF,oFAAoF;QACpF,uFAAuF;QACvF,gCAAgC;QAChC,MAAM6B,QAAQxC,MAAMtB,KAAK,CAACiC,MAAM,KAAK,KAAK4B,QAAQ,MAAMvC,MAAMtB,KAAK,CAACiC,MAAM;QAE1E,MAAM8B,aAAaD,QAAQ3C,WAAWR,IAAIW,MAAMtB,KAAK,IAAIqB,mBAAmBV,IAAIW;QAChFA,MAAM0C,YAAY,GAAGD,cAAcJ;IACrC;AACF,EAAE"}

package/dist/esm/features/rank.js

@@ -8,7 +8,14 @@ export const rank = {
         const columns = db.prepare('PRAGMA table_info(frontmatter)').all();
         if (!columns.some((c)=>c.name === '_rank')) db.exec('ALTER TABLE frontmatter ADD COLUMN "_rank" REAL');
     },
-    afterReconcile (db) {
+    afterReconcile (db, delta) {
+        // links.afterReconcile runs first (registry order) and sets delta.linksChanged;
+        // `undefined` (links off, or no opinion) recomputes, same as always -- only an
+        // explicit `false` allows a skip. Rank also depends on the node set (every note is a
+        // PageRank node, edges or not), so any add or vanish recomputes even with edges
+        // untouched -- otherwise a new linkless note keeps _rank NULL and every other rank
+        // stays normalized over the old count.
+        if (delta.linksChanged === false && delta.added.length === 0 && delta.vanished.length === 0) return;
         const nodes = db.prepare('SELECT "path" FROM frontmatter').all().map((r)=>r.path);
         const ranks = pagerank(nodes, linkEdges(db));
         const update = db.prepare('UPDATE frontmatter SET "_rank" = ? WHERE "path" = ?');

package/dist/esm/features/rank.js.map

@@ -1 +1 @@
-{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/rank.ts"],"sourcesContent":["import { pagerank } from '../graph.ts';\nimport { linkEdges } from './links.ts';\nimport type { Feature } from './types.ts';\n\n// frontmatter._rank: PageRank over resolved links -- a static \"how load-bearing is this\n// note\" prior. Depends on links (enforced in config.featureEnabled).\n\nexport const rank: Feature = {\n  name: 'rank',\n  schema(db) {\n    const columns = db.prepare('PRAGMA table_info(frontmatter)').all() as Array<{ name: string }>;\n    if (!columns.some((c) => c.name === '_rank')) db.exec('ALTER TABLE frontmatter ADD COLUMN \"_rank\" REAL');\n  },\n  afterReconcile(db) {\n    const nodes = (db.prepare('SELECT \"path\" FROM frontmatter').all() as Array<{ path: string }>).map((r) => r.path);\n    const ranks = pagerank(nodes, linkEdges(db));\n    const update = db.prepare('UPDATE frontmatter SET \"_rank\" = ? WHERE \"path\" = ?');\n    for (const [path, value] of ranks) update.run(value, path);\n  },\n};\n"],"names":["pagerank","linkEdges","rank","name","schema","db","columns","prepare","all","some","c","exec","afterReconcile","nodes","map","r","path","ranks","update","value","run"],"mappings":"AAAA,SAASA,QAAQ,QAAQ,cAAc;AACvC,SAASC,SAAS,QAAQ,aAAa;AAGvC,wFAAwF;AACxF,qEAAqE;AAErE,OAAO,MAAMC,OAAgB;IAC3BC,MAAM;IACNC,QAAOC,EAAE;QACP,MAAMC,UAAUD,GAAGE,OAAO,CAAC,kCAAkCC,GAAG;QAChE,IAAI,CAACF,QAAQG,IAAI,CAAC,CAACC,IAAMA,EAAEP,IAAI,KAAK,UAAUE,GAAGM,IAAI,CAAC;IACxD;IACAC,gBAAeP,EAAE;QACf,MAAMQ,QAAQ,AAACR,GAAGE,OAAO,CAAC,kCAAkCC,GAAG,GAA+BM,GAAG,CAAC,CAACC,IAAMA,EAAEC,IAAI;QAC/G,MAAMC,QAAQjB,SAASa,OAAOZ,UAAUI;QACxC,MAAMa,SAASb,GAAGE,OAAO,CAAC;QAC1B,KAAK,MAAM,CAACS,MAAMG,MAAM,IAAIF,MAAOC,OAAOE,GAAG,CAACD,OAAOH;IACvD;AACF,EAAE"}
+{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/rank.ts"],"sourcesContent":["import { pagerank } from '../graph.ts';\nimport { linkEdges } from './links.ts';\nimport type { Feature } from './types.ts';\n\n// frontmatter._rank: PageRank over resolved links -- a static \"how load-bearing is this\n// note\" prior. Depends on links (enforced in config.featureEnabled).\n\nexport const rank: Feature = {\n  name: 'rank',\n  schema(db) {\n    const columns = db.prepare('PRAGMA table_info(frontmatter)').all() as Array<{ name: string }>;\n    if (!columns.some((c) => c.name === '_rank')) db.exec('ALTER TABLE frontmatter ADD COLUMN \"_rank\" REAL');\n  },\n  afterReconcile(db, delta) {\n    // links.afterReconcile runs first (registry order) and sets delta.linksChanged;\n    // `undefined` (links off, or no opinion) recomputes, same as always -- only an\n    // explicit `false` allows a skip. Rank also depends on the node set (every note is a\n    // PageRank node, edges or not), so any add or vanish recomputes even with edges\n    // untouched -- otherwise a new linkless note keeps _rank NULL and every other rank\n    // stays normalized over the old count.\n    if (delta.linksChanged === false && delta.added.length === 0 && delta.vanished.length === 0) return;\n    const nodes = (db.prepare('SELECT \"path\" FROM frontmatter').all() as Array<{ path: string }>).map((r) => r.path);\n    const ranks = pagerank(nodes, linkEdges(db));\n    const update = db.prepare('UPDATE frontmatter SET \"_rank\" = ? WHERE \"path\" = ?');\n    for (const [path, value] of ranks) update.run(value, path);\n  },\n};\n"],"names":["pagerank","linkEdges","rank","name","schema","db","columns","prepare","all","some","c","exec","afterReconcile","delta","linksChanged","added","length","vanished","nodes","map","r","path","ranks","update","value","run"],"mappings":"AAAA,SAASA,QAAQ,QAAQ,cAAc;AACvC,SAASC,SAAS,QAAQ,aAAa;AAGvC,wFAAwF;AACxF,qEAAqE;AAErE,OAAO,MAAMC,OAAgB;IAC3BC,MAAM;IACNC,QAAOC,EAAE;QACP,MAAMC,UAAUD,GAAGE,OAAO,CAAC,kCAAkCC,GAAG;QAChE,IAAI,CAACF,QAAQG,IAAI,CAAC,CAACC,IAAMA,EAAEP,IAAI,KAAK,UAAUE,GAAGM,IAAI,CAAC;IACxD;IACAC,gBAAeP,EAAE,EAAEQ,KAAK;QACtB,gFAAgF;QAChF,+EAA+E;QAC/E,qFAAqF;QACrF,gFAAgF;QAChF,mFAAmF;QACnF,uCAAuC;QACvC,IAAIA,MAAMC,YAAY,KAAK,SAASD,MAAME,KAAK,CAACC,MAAM,KAAK,KAAKH,MAAMI,QAAQ,CAACD,MAAM,KAAK,GAAG;QAC7F,MAAME,QAAQ,AAACb,GAAGE,OAAO,CAAC,kCAAkCC,GAAG,GAA+BW,GAAG,CAAC,CAACC,IAAMA,EAAEC,IAAI;QAC/G,MAAMC,QAAQtB,SAASkB,OAAOjB,UAAUI;QACxC,MAAMkB,SAASlB,GAAGE,OAAO,CAAC;QAC1B,KAAK,MAAM,CAACc,MAAMG,MAAM,IAAIF,MAAOC,OAAOE,GAAG,CAACD,OAAOH;IACvD;AACF,EAAE"}

package/dist/esm/features/types.d.ts

@@ -1,6 +1,13 @@
 import type { DatabaseSync } from 'node:sqlite';
-import type { FeatureName } from '../config.js';
+import type { Config, FeatureName } from '../config.js';
 import type { FileStat } from '../scan.js';
+export interface ReconcileDelta {
+    files: FileStat[];
+    reparsed: string[];
+    added: string[];
+    vanished: string[];
+    linksChanged?: boolean;
+}
 export interface Feature {
     name: FeatureName;
     schema(db: DatabaseSync): void;
@@ -8,7 +15,8 @@ export interface Feature {
         title: string;
         summary: string;
     }): unknown;
-    remove?(db: DatabaseSync, path: string): void;
-    store?(db: DatabaseSync, path: string, extracted: unknown): void;
-    afterReconcile?(db: DatabaseSync, files: FileStat[]): void;
+    remove?(db: DatabaseSync, path: string, delta: ReconcileDelta): void;
+    store?(db: DatabaseSync, path: string, extracted: unknown, delta: ReconcileDelta): void;
+    afterReconcile?(db: DatabaseSync, delta: ReconcileDelta): void;
+    enabledForFile?(cfg: Config, file: FileStat): boolean;
 }

package/dist/esm/features/types.js.map

@@ -1 +1 @@
-{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/types.ts"],"sourcesContent":["import type { DatabaseSync } from 'node:sqlite';\nimport type { FeatureName } from '../config.ts';\nimport type { FileStat } from '../scan.ts';\n\n// A feature owns its schema, per-file extraction/storage, and any whole-tree pass.\n// Adding a feature = one module here + one entry in the registry (index.ts).\nexport interface Feature {\n  name: FeatureName;\n  schema(db: DatabaseSync): void;\n  // Pure, per file: raw is the full file, body the prose after frontmatter,\n  // search the normalized title/summary strings.\n  extract?(raw: string, body: string, search?: { title: string; summary: string }): unknown;\n  remove?(db: DatabaseSync, path: string): void;\n  store?(db: DatabaseSync, path: string, extracted: unknown): void;\n  // After all rows are current, inside the reconcile transaction (resolution, rank).\n  afterReconcile?(db: DatabaseSync, files: FileStat[]): void;\n}\n"],"names":[],"mappings":"AAIA,mFAAmF;AACnF,6EAA6E;AAC7E,WAUC"}
+{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/features/types.ts"],"sourcesContent":["import type { DatabaseSync } from 'node:sqlite';\nimport type { Config, FeatureName } from '../config.ts';\nimport type { FileStat } from '../scan.ts';\n\n// What changed this reconcile, so afterReconcile hooks can act on the delta instead of\n// re-reading the whole tree. `reparsed`/`added`/`vanished` are relPaths; `added` is the\n// subset of `reparsed` that are brand-new files (not just modified). One delta object per\n// reconcile: hooks communicate through it (links sets `linksChanged` for rank), and any\n// state keyed on it dies with the reconcile -- including on ROLLBACK.\nexport interface ReconcileDelta {\n  files: FileStat[];\n  reparsed: string[];\n  added: string[];\n  vanished: string[];\n  // Set by the links feature: the resolved edge set may differ from before this reconcile.\n  linksChanged?: boolean;\n}\n\n// A feature owns its schema, per-file extraction/storage, and any whole-tree pass.\n// Adding a feature = one module here + one entry in the registry (index.ts).\nexport interface Feature {\n  name: FeatureName;\n  schema(db: DatabaseSync): void;\n  // Pure, per file: raw is the full file, body the prose after frontmatter,\n  // search the normalized title/summary strings.\n  extract?(raw: string, body: string, search?: { title: string; summary: string }): unknown;\n  remove?(db: DatabaseSync, path: string, delta: ReconcileDelta): void;\n  store?(db: DatabaseSync, path: string, extracted: unknown, delta: ReconcileDelta): void;\n  // After all rows are current, inside the reconcile transaction (resolution, rank).\n  // Hooks run in registry order and communicate through fields on `delta`.\n  afterReconcile?(db: DatabaseSync, delta: ReconcileDelta): void;\n  // Per-file opt-out (currently only embed, via FileStat.embed): absent means the feature\n  // applies to every file it's otherwise on for. Checked per file before extract/store so a\n  // doc covered only by semantic-false presets gets no rows for it at all.\n  enabledForFile?(cfg: Config, file: FileStat): boolean;\n}\n"],"names":[],"mappings":"AAkBA,mFAAmF;AACnF,6EAA6E;AAC7E,WAeC"}

package/dist/esm/index.d.ts

@@ -1,4 +1,6 @@
-export type { Config, EmbedConfig, FeatureName, ResolvedConfig } from './config.js';
+export type { Peek, PresetCoverage, SearchOptions, TreeMap } from './commands.js';
+export { mapTree, peek, presetCoverage, search } from './commands.js';
+export type { Config, EmbedConfig, FeatureName, Preset, ResolvedConfig, SavedQuery, SavedSearch } from './config.js';
 export { CONFIG_FILENAME, initConfig, loadConfig, migrateConfig, STATE_DIR, SUPPORTED_CONFIG_VERSION } from './config.js';
 export type { OpenResult } from './db.js';
 export { open, rebuild } from './db.js';
@@ -6,7 +8,5 @@ export type { SenseErrorCode } from './errors.js';
 export { SenseError } from './errors.js';
 export type { Row } from './output.js';
 export { printRows } from './output.js';
-export type { FindOptions, Peek, TreeMap } from './verbs.js';
-export { find, mapTree, peek } from './verbs.js';
 export type { WatchEvent, WatchOptions } from './watch.js';
 export { runWatch } from './watch.js';

package/dist/esm/index.js

@@ -1,8 +1,8 @@
 // Public library API. Deliberately small: every export is a stability promise;
 // internals (feature registry, graph, scan, meta) stay module-private.
+export { mapTree, peek, presetCoverage, search } from './commands.js';
 export { CONFIG_FILENAME, initConfig, loadConfig, migrateConfig, STATE_DIR, SUPPORTED_CONFIG_VERSION } from './config.js';
 export { open, rebuild } from './db.js';
 export { SenseError } from './errors.js';
 export { printRows } from './output.js';
-export { find, mapTree, peek } from './verbs.js';
 export { runWatch } from './watch.js';

package/dist/esm/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/index.ts"],"sourcesContent":["// Public library API. Deliberately small: every export is a stability promise;\n// internals (feature registry, graph, scan, meta) stay module-private.\n\nexport type { Config, EmbedConfig, FeatureName, ResolvedConfig } from './config.ts';\nexport { CONFIG_FILENAME, initConfig, loadConfig, migrateConfig, STATE_DIR, SUPPORTED_CONFIG_VERSION } from './config.ts';\n\nexport type { OpenResult } from './db.ts';\nexport { open, rebuild } from './db.ts';\n\nexport type { SenseErrorCode } from './errors.ts';\nexport { SenseError } from './errors.ts';\n\nexport type { Row } from './output.ts';\nexport { printRows } from './output.ts';\n\nexport type { FindOptions, Peek, TreeMap } from './verbs.ts';\nexport { find, mapTree, peek } from './verbs.ts';\n\nexport type { WatchEvent, WatchOptions } from './watch.ts';\nexport { runWatch } from './watch.ts';\n"],"names":["CONFIG_FILENAME","initConfig","loadConfig","migrateConfig","STATE_DIR","SUPPORTED_CONFIG_VERSION","open","rebuild","SenseError","printRows","find","mapTree","peek","runWatch"],"mappings":"AAAA,+EAA+E;AAC/E,uEAAuE;AAGvE,SAASA,eAAe,EAAEC,UAAU,EAAEC,UAAU,EAAEC,aAAa,EAAEC,SAAS,EAAEC,wBAAwB,QAAQ,cAAc;AAG1H,SAASC,IAAI,EAAEC,OAAO,QAAQ,UAAU;AAGxC,SAASC,UAAU,QAAQ,cAAc;AAGzC,SAASC,SAAS,QAAQ,cAAc;AAGxC,SAASC,IAAI,EAAEC,OAAO,EAAEC,IAAI,QAAQ,aAAa;AAGjD,SAASC,QAAQ,QAAQ,aAAa"}
+{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/index.ts"],"sourcesContent":["// Public library API. Deliberately small: every export is a stability promise;\n// internals (feature registry, graph, scan, meta) stay module-private.\n\nexport type { Peek, PresetCoverage, SearchOptions, TreeMap } from './commands.ts';\nexport { mapTree, peek, presetCoverage, search } from './commands.ts';\nexport type { Config, EmbedConfig, FeatureName, Preset, ResolvedConfig, SavedQuery, SavedSearch } from './config.ts';\nexport { CONFIG_FILENAME, initConfig, loadConfig, migrateConfig, STATE_DIR, SUPPORTED_CONFIG_VERSION } from './config.ts';\nexport type { OpenResult } from './db.ts';\nexport { open, rebuild } from './db.ts';\nexport type { SenseErrorCode } from './errors.ts';\nexport { SenseError } from './errors.ts';\nexport type { Row } from './output.ts';\nexport { printRows } from './output.ts';\n\nexport type { WatchEvent, WatchOptions } from './watch.ts';\nexport { runWatch } from './watch.ts';\n"],"names":["mapTree","peek","presetCoverage","search","CONFIG_FILENAME","initConfig","loadConfig","migrateConfig","STATE_DIR","SUPPORTED_CONFIG_VERSION","open","rebuild","SenseError","printRows","runWatch"],"mappings":"AAAA,+EAA+E;AAC/E,uEAAuE;AAGvE,SAASA,OAAO,EAAEC,IAAI,EAAEC,cAAc,EAAEC,MAAM,QAAQ,gBAAgB;AAEtE,SAASC,eAAe,EAAEC,UAAU,EAAEC,UAAU,EAAEC,aAAa,EAAEC,SAAS,EAAEC,wBAAwB,QAAQ,cAAc;AAE1H,SAASC,IAAI,EAAEC,OAAO,QAAQ,UAAU;AAExC,SAASC,UAAU,QAAQ,cAAc;AAEzC,SAASC,SAAS,QAAQ,cAAc;AAGxC,SAASC,QAAQ,QAAQ,aAAa"}

package/dist/esm/output.d.ts

@@ -1,5 +1,15 @@
 export type Row = Record<string, unknown>;
 export declare function printRows(rows: Row[], format: 'table' | 'json'): void;
+export declare function featureOffHint(name: string): string;
+export declare function featuresLine(features: {
+    on: string[];
+    off: string[];
+}): string;
+export declare function presetsLines(presets: Array<{
+    name: string;
+    files: number;
+    embedded: number;
+}>): string[];
 export declare function renderMap(result: {
     docs: {
         count: number;
@@ -11,6 +21,11 @@ export declare function renderMap(result: {
         on: string[];
         off: string[];
     };
+    presets: Array<{
+        name: string;
+        files: number;
+        embedded: number;
+    }>;
     hubs: Row[];
     recent: Row[];
 }): string;
@@ -22,6 +37,7 @@ export declare function renderPeek(result: {
     outbound: string[];
     backlinks: string[];
     unresolved: string[];
+    sectionsTotal: number;
     outboundTotal: number;
     backlinksTotal: number;
     unresolvedTotal: number;

package/dist/esm/output.js

@@ -47,20 +47,34 @@ function renderRows(rows, format, width = process.stdout.columns) {
         ...cells.map(formatRow)
     ].join('\n');
 }
-// Text renderers for the layer verbs; cli.ts prints what these return.
-function featuresLine(features) {
-    const off = features.off.length > 0 ? ` · off: ${features.off.map((name)=>`${name} (features.${name})`).join(', ')}` : '';
+// Text renderers for the top-level commands; cli.ts prints what these return.
+// `embed` is not a features-block key (it's derived from presets), so its off hint reads
+// differently from links/sections/rank's plain `features.<name>`.
+export function featureOffHint(name) {
+    return name === 'embed' ? 'no preset has semantic on' : `features.${name}`;
+}
+export function featuresLine(features) {
+    const off = features.off.length > 0 ? ` · off: ${features.off.map((name)=>`${name} (${featureOffHint(name)})`).join(', ')}` : '';
     return `features: ${features.on.join(', ')}${off}`;
 }
+export function presetsLines(presets) {
+    return [
+        'presets:',
+        ...presets.map((p)=>`  ${p.name}: ${p.files} file(s), ${p.embedded} embedded`)
+    ];
+}
 export function renderMap(result) {
     const parts = [
         `docs: ${result.docs.count} (${Math.round(result.docs.bytes / 1024)} KB)`,
-        `${featuresLine(result.features)}\n`,
+        featuresLine(result.features),
+        '',
+        ...presetsLines(result.presets),
+        '',
         renderRows(result.fields, 'table')
     ];
     if (result.fieldsTotal > result.fields.length) parts.push(`(+${result.fieldsTotal - result.fields.length} more fields)`);
     if (result.hubs.length > 0) parts.push('\nhubs (by link rank):', renderRows(result.hubs, 'table'));
-    else if (result.features.off.includes('rank')) parts.push('\nhubs: off (features.rank)');
+    else if (result.features.off.includes('rank')) parts.push(`\nhubs: off (${featureOffHint('rank')})`);
     parts.push('\nrecent:', renderRows(result.recent, 'table'));
     return parts.join('\n');
 }
@@ -72,6 +86,7 @@ export function renderPeek(result) {
     if (result.sections.length > 0) {
         lines.push('', 'sections:');
         for (const s of result.sections)lines.push(`  ${'#'.repeat(s.level)} ${s.heading}  [L${s.start_line}-${s.end_line}, ~${s.tokens}t]`);
+        if (result.sectionsTotal > result.sections.length) lines.push(`  (+${result.sectionsTotal - result.sections.length} more sections -- sections table has all of them)`);
     } else if (result.off.includes('sections')) lines.push('', 'sections: off (features.sections)');
     if (result.off.includes('links')) {
         lines.push('', 'links: off (features.links)');