pi-mega-compact 0.21.8 → 0.21.10

package/dist/extensions/mega-runtime/dashboard-snapshot.js

@@ -128,6 +128,7 @@ export function buildDashboardSnapshot(ctx) {
             ctxFastGate: ctx.diagCtxFastGate,
             liveTrimFires: ctx.diagLiveTrimFires,
             liveTrimReplays: ctx.diagLiveTrimReplays,
+            headroomTrips: ctx.diagCtxHeadroomTrip,
         },
         retries: {
             errorRetryCount: ctx.errorRetryCount,

package/dist/extensions/mega-runtime/runtime-instrumentation.js

@@ -23,6 +23,7 @@ export class RuntimeInstrumentation {
     diagCtxCutNull = 0; // computeLiveTrimCut returned null (anchor/boundary)
     diagCtxThrown = 0; // live-trim try threw (caught)
     diagCtxOutputErrorTrip = 0; // Phase H: output-error catch tripped a forced compaction
+    diagCtxHeadroomTrip = 0; // v0.21.9: output-headroom gate tripped a pre-overflow compaction
     // Context health instrumentation (v0.12): rolling ring buffers for
     // drift detection + cache poison Layer 1 hash baseline.
     recentTurnEmbeddings = [];

package/dist/extensions/mega-runtime/runtime-snapshot.js

@@ -94,6 +94,7 @@ export function snapshotImpl(self, ctx) {
         diagCtxFastGate: self.diagCtxFastGate,
         diagLiveTrimFires: self.diagLiveTrimFires,
         diagLiveTrimReplays: self.diagLiveTrimReplays,
+        diagCtxHeadroomTrip: self.diagCtxHeadroomTrip,
         errorRetryCount: self.rt.errorRetryCount,
         consecutiveErrors: self.rt.consecutiveErrors,
         ERROR_RETRY_MAX_CONSECUTIVE: self.config.maxConsecutiveErrors,

package/dist/src/config/dedup.js

@@ -63,6 +63,9 @@ export function loadDedupConfig() {
         L2_COSINE_CODE: envNumOrNull("MEGACOMPACT_L2_THRESHOLD_CODE"),
         L2_COSINE_PROSE: envNumOrNull("MEGACOMPACT_L2_THRESHOLD_PROSE"),
         L1_JACCARD: envNum("MEGACOMPACT_L1_JACCARD", 0.8),
+        DEDUP_DEGENERATE_GUARD: envBool("MEGACOMPACT_DEDUP_DEGENERATE_GUARD", true),
+        DEDUP_DEGEN_MIN_TOKENS: envNum("MEGACOMPACT_DEDUP_DEGEN_MIN_TOKENS", 48),
+        DEDUP_DEGEN_MIN_PCT: envNum("MEGACOMPACT_DEDUP_DEGEN_MIN_PCT", 0.005),
         DEDUP_SIM: envNum("MEGACOMPACT_DEDUP_SIM", 0.9),
         MMR_LAMBDA: envNum("MEGACOMPACT_MMR_LAMBDA", 0.5),
         SEMDEDUP_COSINE: envNum("MEGACOMPACT_SEMDEDUP_COSINE", 0.95),

package/dist/src/dedup/degenerate.js

@@ -0,0 +1,68 @@
+/**
+ * The effective token floor for a checkpoint: the larger of the absolute floor
+ * and `MIN_PCT × originalTokenEstimate`.
+ *
+ * A missing / zero / non-finite `originalTokenEstimate` contributes nothing, so
+ * the absolute floor applies alone — direct add() callers and pre-v0.4 rows that
+ * never recorded the original region size are judged on absolute size only,
+ * never accidentally deemed degenerate by a 0-valued percentage term.
+ */
+export function degenerateFloor(subject, tunables) {
+    const orig = subject.originalTokenEstimate;
+    const relative = typeof orig === "number" && Number.isFinite(orig) && orig > 0
+        ? orig * tunables.DEDUP_DEGEN_MIN_PCT
+        : 0;
+    return Math.max(tunables.DEDUP_DEGEN_MIN_TOKENS, relative);
+}
+/**
+ * Is this stored checkpoint a degenerate (content-free) summary?
+ *
+ * Calibration against the incident data:
+ *  - skeleton: tokenEstimate 34, original ≈19166 → 34 < max(48, 95.8) → TRUE
+ *  - normal:   tokenEstimate 2000, original 70000 → 2000 > max(48, 350) → FALSE
+ *
+ * The relative term is what makes this scale: a 34-token summary of a 900-token
+ * region is a legitimate 26× compression, while the same 34 tokens standing in
+ * for 19k is a skeleton.
+ */
+export function isDegenerateCheckpoint(subject, tunables) {
+    const tokens = subject.tokenEstimate ?? 0;
+    return tokens < degenerateFloor(subject, tunables);
+}
+/**
+ * Should an L1/L2 match be DECLINED because it would collapse richer incoming
+ * content onto a degenerate stored checkpoint?
+ *
+ * Returns true only when all four hold:
+ *   1. the umbrella flag is ON,
+ *   2. the matched (stored) checkpoint is degenerate,
+ *   3. the candidate is strictly richer than the match,
+ *   4. the candidate's content is not byte-identical to the match's.
+ *
+ * Condition 3 uses a strict `>`: equal-size skeletons collapsing is harmless and
+ * keeps the store from growing one row per compaction while the summarizer is
+ * broken. Only a genuine improvement is worth declining a collapse for.
+ *
+ * Condition 4 is a CORRECTNESS requirement, not a refinement. `context_chunks`
+ * carries a partial UNIQUE index on (session_id, content_hash) (schema/core.ts
+ * QA #1), so declining a match whose content hash already exists would fall
+ * through to an INSERT that throws — inside add(), which sits on the agent loop.
+ * It is also the semantically right call: identical bytes are the SAME region,
+ * so re-storing them adds no information and heals nothing. Only L0 may own the
+ * exact-match case; the guard exists for fuzzy matches on genuinely different
+ * text, which is exactly the incident's shape (each compaction produced a
+ * *similar but distinct* skeleton).
+ */
+export function shouldSkipDegenerateMatch(matched, candidate, tunables) {
+    if (!tunables.DEDUP_DEGENERATE_GUARD)
+        return false;
+    if (!isDegenerateCheckpoint(matched, tunables))
+        return false;
+    if ((candidate.tokenEstimate ?? 0) <= (matched.tokenEstimate ?? 0))
+        return false;
+    // Byte-identical content → not a healing opportunity (and would violate the
+    // UNIQUE index). Compared only when both hashes are known.
+    const a = candidate.contentHash;
+    const b = matched.contentHash;
+    return !(a !== undefined && b !== undefined && a === b);
+}

package/dist/src/extractive-salvage.js

@@ -0,0 +1,195 @@
+/**
+ * extractive-salvage.ts — file-path policy + skeleton salvage for extractive.ts.
+ *
+ * Extracted from extractive.ts (A1/A2 sprint) to keep that file under the
+ * 300-line src/ soft limit. Pure functions only — no I/O, no logging.
+ *
+ * DESIGN (A1): blocklist, not allowlist. The old `INTERESTING_EXT` allowlist
+ * (rs/ts/tsx/js/json/md) silently produced content-free summaries for every
+ * other language — a .go/.py/.c project got a 34-token skeleton. An allowlist
+ * has to be *right* about ~40 ecosystems to be useful and fails closed (drops
+ * real work) when it is wrong; a NOISE blocklist only has to be right about the
+ * small, stable set of binary/generated/asset extensions and fails open (an
+ * unknown extension is surfaced, which is the safe direction for a summary).
+ */
+// ---- File path policy ------------------------------------------------------
+/**
+ * Generic extension capture: any 1–6 char ALPHABETIC extension. Path
+ * character class is unchanged from the original FILE_PATH_RE so existing
+ * matching behaviour (quotes/backticks/whitespace as delimiters) is preserved.
+ * Alphabetic-only is deliberate: every real source/config extension is alpha
+ * (c..tsx, tsconfig.json), while an alphanumeric class matches version strings
+ * ("GLM-4.7", "v0.21.9", "Node 18.2") as "files" and spams Key files.
+ */
+export const FILE_PATH_RE = /(?:^|\s)([^\s"`']+\.([A-Za-z]{1,6}))\b/g;
+/** Same policy, single-match, for inferCurrentWork (also excludes ':'). */
+export const CURRENT_WORK_PATH_RE = /(?:^|\s)([^\s"`':]+\.([A-Za-z]{1,6}))\b/m;
+/**
+ * Binary / generated / asset / vendored extensions that carry no summary value.
+ * Deliberately small and stable. Note `md`, `json`, `toml`, `yaml`, `sql`, `css`
+ * and `html` are NOT noise — they are hand-edited source in most repos.
+ */
+export const NOISE_EXT = new Set([
+    // lockfiles & logs
+    "lock", "log", "sum",
+    // images & media
+    "png", "jpg", "jpeg", "gif", "webp", "svg", "ico", "bmp", "tiff",
+    "mp3", "mp4", "mov", "wav", "webm", "avi",
+    // fonts
+    "woff", "woff2", "ttf", "otf", "eot",
+    // archives & binaries
+    "zip", "gz", "tgz", "bz2", "xz", "7z", "rar", "tar",
+    "exe", "dll", "so", "dylib", "bin", "o", "a", "obj", "class", "pyc", "pyo",
+    "wasm", "node", "jar", "war", "deb", "rpm", "dmg", "iso", "img",
+    // generated / build artifacts
+    "map", "min", "lockb", "snap", "cache", "tmp", "temp", "swp", "bak", "orig",
+    // data blobs & databases
+    "pdf", "db", "sqlite", "sqlite3", "pack", "idx", "pem", "key", "crt",
+]);
+/** Directory fragments whose files are never "key files" for a summary. */
+const NOISE_DIR_RE = /(?:^|\/)(?:node_modules|\.git|dist|build|coverage|vendor|target|__pycache__|\.venv|venv)(?:\/|$)/;
+/**
+ * TLD-shaped extensions: `example.com`, `github.com`, `npm.cmd` are domains and
+ * launchers, not code. (QA lens 1 finding, 2026-08-19.) Conservative list only —
+ * no real source extension lives here (.go/.rs/.ts stay interesting).
+ */
+const TLD_EXT = new Set([
+    "com", "org", "net", "io", "gov", "edu", "biz", "info", "dev", "app",
+    "page", "xyz", "site", "online", "cloud", "me", "co", "us", "uk", "de",
+    "fr", "jp", "cn", "nl", "se", "eu", "int", "mil", "cmd",
+]);
+/** True when a matched path is worth surfacing in a summary. */
+export function isInterestingPath(filePath, ext) {
+    const lowerExt = ext.toLowerCase();
+    if (NOISE_EXT.has(lowerExt))
+        return false;
+    if (TLD_EXT.has(lowerExt))
+        return false; // bare domains, not code
+    if (/^(?:https?|ftp):\/\//i.test(filePath) || filePath.toLowerCase().startsWith("www."))
+        return false;
+    if (NOISE_DIR_RE.test(filePath))
+        return false;
+    // `app.min.js` / `bundle.min.css` style double extensions.
+    if (/\.min\.[A-Za-z0-9]{1,6}$/.test(filePath))
+        return false;
+    if (/\.map$/.test(filePath))
+        return false;
+    // Prose abbreviations ("e.g", "i.e", "U.S"): a single-char base with no slash
+    // and no digit is never a filename. ("a.c" is rare collateral; "q1.py" and
+    // "src/a.ts" survive — digit / slash both exempt.)
+    const base = filePath.slice(0, filePath.lastIndexOf("."));
+    if (!filePath.includes("/") && !/\d/.test(filePath) && base.length <= 1)
+        return false;
+    return true;
+}
+// ---- Path collection -------------------------------------------------------
+const MAX_KEY_FILES = 5;
+const MAX_FILES = 10;
+const FRESHNESS_WINDOW = 10;
+/** All interesting file paths mentioned in a blob of text. */
+export function extractFilePaths(text) {
+    const paths = [];
+    for (const m of text.matchAll(FILE_PATH_RE)) {
+        if (isInterestingPath(m[1], m[2]))
+            paths.push(m[1]);
+    }
+    return paths;
+}
+/** Most-mentioned paths within the recency window (behaviour unchanged). */
+export function collectKeyFiles(messages) {
+    const recent = messages.slice(-FRESHNESS_WINDOW);
+    const pathFreq = new Map();
+    for (const m of recent) {
+        for (const p of extractFilePaths(m.text)) {
+            pathFreq.set(p, (pathFreq.get(p) ?? 0) + 1);
+        }
+    }
+    return [...pathFreq.entries()]
+        .sort((a, b) => b[1] - a[1])
+        .slice(0, MAX_KEY_FILES)
+        .map(([p]) => p);
+}
+/** Paths written/edited by tools (extension-agnostic; behaviour unchanged). */
+export function extractFilesModified(tools) {
+    const files = new Set();
+    for (const m of tools) {
+        if (!m.toolName)
+            continue;
+        const name = m.toolName.toLowerCase();
+        if (name === "write" || name === "edit" || name === "notebookedit") {
+            const input = m.input ?? m.text;
+            const pathMatch = input.match(/["']?(\/[^\s"']+\.\w+)["']?/);
+            if (pathMatch)
+                files.add(pathMatch[1]);
+        }
+        if (name === "bash") {
+            const cmd = m.input ?? m.text;
+            if (cmd.includes("git add") || cmd.includes("git commit") || cmd.includes("git diff")) {
+                for (const p of extractFilePaths(cmd))
+                    files.add(p);
+            }
+        }
+    }
+    return [...files].slice(0, MAX_FILES);
+}
+// ---- Placeholder user turns (A2b) ------------------------------------------
+/**
+ * Content-free user turns. Extremely common in resumed sessions; taking them
+ * verbatim as "User requests" is what produced the three "• resume" bullets.
+ */
+const PLACEHOLDER_RE = /^(?:resume|continue|go on|go ahead|proceed|next|yes|yeah|yep|y|ok|okay|k|sure|thanks|thank you|ty|done|please continue|carry on)\W*$/i;
+export function isPlaceholderRequest(text) {
+    return PLACEHOLDER_RE.test(text.trim());
+}
+// ---- Skeleton salvage (A2c) ------------------------------------------------
+const MAX_SALVAGE_LINES = 5;
+const SALVAGE_LINE_LEN = 120;
+/**
+ * A "skeleton" summary is the scope line and nothing else — no files, no current
+ * work, no decisions, no pending items, and no *substantive* user request. That
+ * is ~34 tokens of zero information and is what breaks a resumed session.
+ *
+ * NOTE: `recentUser` is treated as empty when it holds only placeholders. The
+ * incident summary DID have three "• resume" bullets, so a plain
+ * `recentUser.length === 0` test would never have fired on the very case this
+ * salvage exists for.
+ */
+export function isSkeletonSummary(parts) {
+    const hasRealRequest = parts.recentUser.some((r) => !isPlaceholderRequest(r));
+    return (!hasRealRequest &&
+        parts.keyFiles.length === 0 &&
+        !parts.currentWork &&
+        parts.decisions.length === 0 &&
+        parts.pending.length === 0);
+}
+/**
+ * Last-resort content: the first meaningful line of the most recent assistant
+ * and tool messages. Deterministic (pure scan, newest-first, then re-ordered
+ * oldest-first for reading). Returns [] when there is genuinely nothing.
+ */
+export function buildSalvageDigest(messages) {
+    const out = [];
+    const seen = new Set();
+    for (let i = messages.length - 1; i >= 0 && out.length < MAX_SALVAGE_LINES; i--) {
+        const m = messages[i];
+        if (m.role !== "assistant" && m.role !== "tool")
+            continue;
+        const raw = m.text || m.output || m.input || "";
+        const line = raw
+            .split("\n")
+            .map((l) => l.trim())
+            .find((l) => l.length > 0);
+        if (!line)
+            continue;
+        const label = m.role === "tool" ? `${m.toolName ?? "tool"}: ` : "";
+        const entry = truncateLine(`${label}${line}`, SALVAGE_LINE_LEN);
+        if (seen.has(entry))
+            continue;
+        seen.add(entry);
+        out.push(entry);
+    }
+    return out.reverse();
+}
+function truncateLine(s, maxLen) {
+    return s.length <= maxLen ? s : s.slice(0, maxLen - 1) + "…";
+}

package/dist/src/extractive.js

@@ -8,12 +8,14 @@
  * Deterministic: same messages → same output, every time.
  */
 import { estimateBlockTokens } from "./tokens.js";
+import { CURRENT_WORK_PATH_RE, isInterestingPath, isPlaceholderRequest, isSkeletonSummary, buildSalvageDigest, collectKeyFiles, extractFilesModified, } from "./extractive-salvage.js";
 // ---- Limits ----------------------------------------------------------------
 const MAX_RECENT_USER = 3;
 const MAX_DECISIONS = 5;
-const MAX_FILES = 10;
 const MAX_PENDING = 5;
 const MAX_TOPIC_LINES = 12;
+/** Cap for the merged keyFiles ∪ filesModified "Key files" line (A2a). */
+const MAX_SUMMARY_FILES = 8;
 // ---- Truncation helper -----------------------------------------------------
 function truncate(s, maxLen) {
     if (s.length <= maxLen)
@@ -27,8 +29,22 @@ function truncate(s, maxLen) {
  * Captures: tools used, recent user requests, current work, key files,
  * pending work. Typically 12 lines / ~500 tokens instead of ~70K.
  */
-function buildTopicSummary(messages, tools, recentUser, currentWork, keyFiles, pending) {
+function buildTopicSummary(messages, tools, recentUser, currentWork, keyFiles, pending, filesModified, decisions) {
     const lines = [];
+    // A2a: files captured from write/edit tool inputs are extracted
+    // extension-agnostically but never reached the summary. Fold them in so work
+    // outside the recency window (and outside the path regex) is still reported.
+    // Drop an absolute path when a kept relative path already names the SAME file.
+    // Only MULTI-COMPONENT relative paths fold ("engine/mesh.go" absorbs
+    // "/proj/engine/mesh.go"); a bare basename ("mesh.go") is never folded into an
+    // absolute path, since "/proj/other/x/mesh.go" may be a genuinely different
+    // file (QA lens 1 finding: the naive endsWith dropped different directories).
+    const combined = [...keyFiles, ...filesModified];
+    const allFiles = [...new Set(combined)].filter((p) => {
+        if (!p.startsWith("/"))
+            return true;
+        return !combined.some((r) => r !== p && !r.startsWith("/") && r.includes("/") && p.endsWith("/" + r));
+    }).slice(0, MAX_SUMMARY_FILES);
     // Scope line
     const users = messages.filter((m) => m.role === "user");
     const assistants = messages.filter((m) => m.role === "assistant");
@@ -45,45 +61,59 @@ function buildTopicSummary(messages, tools, recentUser, currentWork, keyFiles, p
     // Current work
     if (currentWork)
         lines.push(`Current work: ${currentWork}`);
-    // Key files
-    if (keyFiles.length)
-        lines.push(`Key files: ${keyFiles.join(", ")}.`);
+    // Key files (keyFiles ∪ filesModified)
+    if (allFiles.length)
+        lines.push(`Key files: ${allFiles.join(", ")}.`);
     // Pending work
     if (pending.length) {
         lines.push("Pending work:");
         for (const p of pending)
             lines.push(`  • ${p}`);
     }
+    // A2c: a scope-line-only summary carries zero information and strands a
+    // resumed session. Salvage the tail of the conversation instead. The line cap
+    // is raised ONLY here: the salvage block is bounded at 5 lines + 1 header, and
+    // a skeleton by definition contributed just the 1 scope line, so the worst
+    // case is 7 lines — still well under the normal 12-line budget.
+    const skeleton = isSkeletonSummary({ recentUser, keyFiles: allFiles, currentWork, decisions, pending });
+    if (skeleton) {
+        const digest = buildSalvageDigest(messages);
+        if (digest.length) {
+            lines.push("Recent activity:");
+            for (const d of digest)
+                lines.push(`  • ${d}`);
+        }
+        return lines.join("\n");
+    }
     // Cap total length
     return lines.slice(0, MAX_TOPIC_LINES).join("\n");
 }
-// ---- File path extraction --------------------------------------------------
-const INTERESTING_EXT = new Set(["rs", "ts", "tsx", "js", "json", "md"]);
-const FILE_PATH_RE = /(?:^|\s)([^\s"`']+\.(rs|ts|tsx|js|json|md|py|sh|sql|toml|yaml|yml|css|html))\b/g;
-function extractFilePaths(text) {
-    const paths = [];
-    for (const m of text.matchAll(FILE_PATH_RE)) {
-        const filePath = m[1];
-        const ext = m[2];
-        const basename = filePath.split("/").pop() ?? filePath;
-        if (basename === "node_modules" || filePath.includes("node_modules/"))
-            continue;
-        if (INTERESTING_EXT.has(ext))
-            paths.push(filePath);
-    }
-    return paths;
-}
-// ---- Recent user requests (existing logic, kept) ---------------------------
+// ---- Recent user requests --------------------------------------------------
+/**
+ * A2b: skip content-free "resume"/"continue" turns and look further back to
+ * fill the quota, so a resumed session surfaces its real requests. Falls back
+ * to the placeholders when EVERY user turn is one (an honest "• resume" beats
+ * an empty section).
+ */
 function collectRecentUserRequests(messages, limit) {
-    const requests = [];
-    for (let i = messages.length - 1; i >= 0 && requests.length < limit; i--) {
-        if (messages[i].role === "user") {
-            let snippet = messages[i].text.split("\n").slice(0, 3).join(" ");
-            snippet = snippet.replace(/^.+\nProcessed\$?\s*/i, "").replace(/\n/g, " ");
-            requests.push(truncate(snippet, 200));
+    const substantive = [];
+    const placeholders = [];
+    for (let i = messages.length - 1; i >= 0 && substantive.length < limit; i--) {
+        if (messages[i].role !== "user")
+            continue;
+        let snippet = messages[i].text.split("\n").slice(0, 3).join(" ");
+        snippet = snippet.replace(/^.+\nProcessed\$?\s*/i, "").replace(/\n/g, " ");
+        const cleaned = truncate(snippet, 200);
+        if (!cleaned.trim())
+            continue;
+        if (isPlaceholderRequest(cleaned)) {
+            if (placeholders.length < limit)
+                placeholders.push(cleaned);
+            continue;
         }
+        substantive.push(cleaned);
     }
-    return requests.reverse();
+    return (substantive.length ? substantive : placeholders).reverse();
 }
 // ---- Pending work (existing logic, kept) -----------------------------------
 const PENDING_WORDS = ["todo", "next", "pending", "follow up", "remaining"];
@@ -106,8 +136,9 @@ function inferCurrentWork(messages) {
         const m = messages[i];
         if (m.role !== "assistant")
             continue;
-        const path = m.text.match(/(?:^|\s)([^\s"`':]+\.(rs|ts|tsx|js|json|md|py|toml|yaml|yml|sql))\b/m);
-        if (path) {
+        // A1: same language-agnostic policy as extractFilePaths.
+        const path = m.text.match(CURRENT_WORK_PATH_RE);
+        if (path && isInterestingPath(path[1], path[2])) {
             const line = m.text.split("\n").slice(0, 2).join(" ");
             return truncate(line, 200);
         }
@@ -144,30 +175,6 @@ function extractDecisions(messages) {
     }
     return [...new Set(decisions)];
 }
-// ---- Files modified --------------------------------------------------------
-function extractFilesModified(tools) {
-    const files = new Set();
-    for (const m of tools) {
-        if (!m.toolName)
-            continue;
-        const name = m.toolName.toLowerCase();
-        if (name === "write" || name === "edit" || name === "notebookedit") {
-            // Extract file path from input payload
-            const input = m.input ?? m.text;
-            const pathMatch = input.match(/["']?(\/[^\s"']+\.\w+)["']?/);
-            if (pathMatch)
-                files.add(pathMatch[1]);
-        }
-        if (name === "bash") {
-            const cmd = m.input ?? m.text;
-            if (cmd.includes("git add") || cmd.includes("git commit") || cmd.includes("git diff")) {
-                for (const p of extractFilePaths(cmd))
-                    files.add(p);
-            }
-        }
-    }
-    return [...files].slice(0, MAX_FILES);
-}
 // ---- Public API ------------------------------------------------------------
 /**
  * Deterministic extractive summary. Same messages → same output, every time.
@@ -192,23 +199,7 @@ export function extractiveSummarize(messages) {
     const pending = inferPendingWork(safe);
     const keyDecisions = extractDecisions(safe);
     const filesModified = extractFilesModified(toolMsgs);
-    const topicSummary = buildTopicSummary(safe, tools, recentUser, currentWork, keyFiles, pending);
+    const topicSummary = buildTopicSummary(safe, tools, recentUser, currentWork, keyFiles, pending, filesModified, keyDecisions);
     const tokenEstimate = estimateBlockTokens(topicSummary);
     return { topicSummary, keyDecisions, nextSteps: pending, filesModified, tokenEstimate };
 }
-// ---- Key files (existing logic from compact.ts, moved here) ----------------
-const MAX_KEY_FILES = 5;
-const FRESHNESS_WINDOW = 10;
-function collectKeyFiles(messages) {
-    const recent = messages.slice(-FRESHNESS_WINDOW);
-    const pathFreq = new Map();
-    for (const m of recent) {
-        for (const p of extractFilePaths(m.text)) {
-            pathFreq.set(p, (pathFreq.get(p) ?? 0) + 1);
-        }
-    }
-    return [...pathFreq.entries()]
-        .sort((a, b) => b[1] - a[1])
-        .slice(0, MAX_KEY_FILES)
-        .map(([p]) => p);
-}

package/dist/src/vector-cortex/dedup-attr/rollup.js

@@ -49,6 +49,11 @@ export function computeDedupTierRollup(events, windowMs, now) {
         const ts = Date.parse(ev.ts);
         if (Number.isNaN(ts) || ts < windowStart || ts > windowEnd)
             continue;
+        // "skipped" (degenerate-match guard declined a collapse) is a decision, but
+        // NOT a tier catch — it is attributed to no tier below. Excluding it from the
+        // denominator keeps l0Share + l1Share + l2Share summing to 1 over the window.
+        if (ev.status === "skipped")
+            continue;
         total += 1;
         switch (ev.tier) {
             case "L0":

package/dist/src/vectorStore/add-degenerate.js

@@ -0,0 +1,25 @@
+import { shouldSkipDegenerateMatch } from "../dedup/degenerate.js";
+/**
+ * Build the decliner for one add() cascade.
+ *
+ * Returns a predicate the L1/L2 call sites use as a one-line guard. When it
+ * returns true it has ALREADY recorded the declined decision (monitoring event +
+ * `skipped` audit line + the live `onTier` detail), so the caller only has to
+ * fall through. When the umbrella flag is off it always returns false and emits
+ * nothing — byte-identical to the pre-guard cascade.
+ */
+export function degenerateDecliner(args) {
+    const { store, input, contentHash, cfg, audit, t0 } = args;
+    const candidate = { ...input, contentHash };
+    return (tier, matched, similarity) => {
+        if (!shouldSkipDegenerateMatch(matched, candidate, cfg))
+            return false;
+        // Reported as `mark_only`: a tier matched but policy declined to collapse —
+        // exactly the existing MARK_ONLY shape, with a distinct reason string so the
+        // dashboard can tell a guard decline from an operator-configured MARK_ONLY.
+        store.record(tier, "mark_only", "degenerateGuard", Date.now() - t0, similarity, matched.checkpointId);
+        audit.skipped(tier, matched.checkpointId, "degenerateGuard", similarity);
+        input.onTier?.({ tier, status: "passed", detail: "degenerateGuard" });
+        return true;
+    };
+}

package/dist/src/vectorStore/add.js

@@ -27,6 +27,7 @@ import { lshBands } from "../dedup/l1-lsh.js";
 import { openBloom, saveBloom } from "../store/bloom.js";
 import { listCheckpoints, nextCheckpointId, upsertCheckpoint, loadSessionState, saveSessionState, upsertMinhashSignature, insertLshBuckets, addTokensSaved, bumpDedupStats, } from "../store/sqlite.js";
 import { computeRegionHash } from "./hash.js";
+import { degenerateDecliner } from "./add-degenerate.js";
 import { runL0Tier, computeSummaryHash } from "./add-l0.js";
 import { findL1Duplicate } from "./add-l1.js";
 import { dedupAuditRecorder } from "./dedup-audit.js";
@@ -67,8 +68,22 @@ export function addCheckpoint(store, input) {
     // Tracks whether a tier matched while in MARK_ONLY (record-but-don't-collapse),
     // and which tier.
     let markOnly = null;
-    // L0 exact-match tier (contentHash / regionHash / summaryHash) — see add-l0.ts.
+    // Content digest for this candidate (also consumed by the L0 tier below).
     const digest = computeContentDigest(input.regionText);
+    // Degenerate-match guard (incident 2026-08-19): declines a fuzzy-tier collapse
+    // onto a content-free skeleton when the incoming region is richer, so the
+    // skeleton stops absorbing every future compaction. Returns true (having
+    // already recorded the decision) ⇒ the caller treats the match as a non-match.
+    // Flag-off ⇒ always false ⇒ byte-identical predecessor. See add-degenerate.ts.
+    const declineDegenerate = degenerateDecliner({
+        store,
+        input,
+        contentHash: digest.contentHash,
+        cfg,
+        audit,
+        t0,
+    });
+    // L0 exact-match tier (contentHash / regionHash / summaryHash) — see add-l0.ts.
     const bloom = openBloom(store.stateDir);
     const summaryHash = computeSummaryHash(input.topicSummary);
     const l0 = runL0Tier({
@@ -94,7 +109,10 @@ export function addCheckpoint(store, input) {
     onTier?.({ tier: "L1", status: "scanning" });
     if (cfg.L1_ENABLED) {
         const l1 = findL1Duplicate(store, sessionId, input.regionText, all);
-        if (l1 && !cfg.MARK_ONLY_L1) {
+        // Guard first: a declined match must not collapse and must not be recorded
+        // as a MARK_ONLY hit either — it is a non-match for the rest of the cascade.
+        const l1Declined = l1 !== undefined && declineDegenerate("L1", l1);
+        if (l1 && !l1Declined && !cfg.MARK_ONLY_L1) {
             l1.timestamp = input.timestamp;
             upsertCheckpoint(l1, store.stateDir);
             bumpDedupStats(true, store.stateDir);
@@ -105,7 +123,7 @@ export function addCheckpoint(store, input) {
             onTier?.({ tier: "L1", status: "deduped", detail: "l1MinHash" });
             return r;
         }
-        if (l1 && cfg.MARK_ONLY_L1)
+        if (l1 && !l1Declined && cfg.MARK_ONLY_L1)
             markOnly = "L1";
     }
     onTier?.({ tier: "L1", status: "passed" });
@@ -131,7 +149,12 @@ export function addCheckpoint(store, input) {
             const sim = cosineSimilarity(embedding, cp.embedding);
             return sim > best.sim ? { checkpoint: cp, sim } : best;
         }, { checkpoint: all[0], sim: -1 });
-        if (!timedOut && nearest.sim >= simThreshold) {
+        // A declined match falls through to the "store a fresh checkpoint" path; the
+        // guard already audited the decision, so the near-miss emit below is skipped.
+        const l2Declined = !timedOut &&
+            nearest.sim >= simThreshold &&
+            declineDegenerate("L2", nearest.checkpoint, nearest.sim);
+        if (!timedOut && !l2Declined && nearest.sim >= simThreshold) {
             if (!cfg.MARK_ONLY_L2) {
                 // Near-identical — update timestamp on existing checkpoint
                 nearest.checkpoint.timestamp = input.timestamp;
@@ -164,7 +187,7 @@ export function addCheckpoint(store, input) {
         // Near-miss: how close did we come to collapsing? Only emitted when the
         // scan actually completed and scored a candidate — a timed-out scan has no
         // honest best to report.
-        if (!timedOut && nearest.sim >= 0) {
+        if (!timedOut && !l2Declined && nearest.sim >= 0) {
             audit.passed("L2", nearest.checkpoint.checkpointId, nearest.sim);
         }
     }

package/dist/src/vectorStore/dedup-audit.js

@@ -69,6 +69,14 @@ export function dedupAuditRecorder(ctx, scope) {
             matchedEntry,
             similarity,
         }),
+        skipped: (tier, matchedEntry, dedupReason, similarity) => emitDedupAudit(ctx, {
+            ...base,
+            tier,
+            status: "skipped",
+            matchedEntry,
+            dedupReason,
+            ...(similarity === undefined ? {} : { similarity }),
+        }),
         stored: (storedEntry, dedupReason, tokenEstimate) => emitDedupAudit(ctx, {
             ...base,
             tier: "new",

package/dist/vector-cortex/dedup-attr/rollup.js

@@ -49,6 +49,11 @@ export function computeDedupTierRollup(events, windowMs, now) {
         const ts = Date.parse(ev.ts);
         if (Number.isNaN(ts) || ts < windowStart || ts > windowEnd)
             continue;
+        // "skipped" (degenerate-match guard declined a collapse) is a decision, but
+        // NOT a tier catch — it is attributed to no tier below. Excluding it from the
+        // denominator keeps l0Share + l1Share + l2Share summing to 1 over the window.
+        if (ev.status === "skipped")
+            continue;
         total += 1;
         switch (ev.tier) {
             case "L0":

package/dist/vectorStore/dedup-audit.js

@@ -69,6 +69,14 @@ export function dedupAuditRecorder(ctx, scope) {
             matchedEntry,
             similarity,
         }),
+        skipped: (tier, matchedEntry, dedupReason, similarity) => emitDedupAudit(ctx, {
+            ...base,
+            tier,
+            status: "skipped",
+            matchedEntry,
+            dedupReason,
+            ...(similarity === undefined ? {} : { similarity }),
+        }),
         stored: (storedEntry, dedupReason, tokenEstimate) => emitDedupAudit(ctx, {
             ...base,
             tier: "new",