@bongos/core 1.17.0

package/scripts/gds/seed-gds-v4-tasks.js

@@ -0,0 +1,624 @@
+#!/usr/bin/env node
+// Seed the 38 net-new GDS-V4 tasks defined in V4 Planning Session #1
+// (2026-06-12, task 937 / V4.R00).
+//
+// Idempotent: skips tasks whose source_ref already exists; dependency wiring
+// is idempotent on (task_id, depends_on_task_id).
+//
+// Companion migration 080_gds_v4_planning_seed.sql handles:
+//   - 6 done_when_criteria seed rows
+//   - versions.GDS-V4 name + done_when text
+//   - 2 existing-task (263, 594) priority + V4.R## title-prefix updates
+//
+// This script handles the 38 net-new task creates, the 5 idea-promotion
+// PATCH calls (ideas 148+175 → V4.R03; 226 → V4.R15; 225 → V4.R16;
+// 178 → V4.R27), and the dependency-edge POSTs.
+//
+// Ranking convention: priority is 1-5 (CHECK constraint). Fine-grained
+// ordering lives in the V4.R## title prefix (zero-padded for sort).
+// Tier mapping: T1/T2=P5, T3/T4=P4, T5=P3, T6=P2, T7=P1.
+//
+// Usage:  node scripts/gds/seed-gds-v4-tasks.js [--dry-run]
+
+const { cliClient, requireSession } = require('./cli-lib');
+
+const DRY = process.argv.includes('--dry-run');
+
+// Each entry: { ref, title, description, kind, discipline, priority,
+//               est_minutes, manual_degree, credits_reward, touches,
+//               deps: ['R##', ...], promoted_idea_id?, also_promote_ideas? }
+// `ref` doubles as source_ref ("gds-v4-r##") and as the V4.R## title prefix.
+
+const TASKS = [
+  // =========================================================================
+  // TIER 1/2 — Roots of both tracks (P5)
+  // =========================================================================
+  {
+    ref: 'R01',
+    title: 'Golden-question recall eval set, mined from session_records',
+    description: `**Criterion:** one-call-recall (also the yardstick for sessions-start-knowing).
+
+Build the 25–40-question golden set that every retrieval phase is measured against. Source the questions from REAL rediscovery moments in the session corpus (e.g. "does the Discord auto-join flow already exist?", "where is recordApplause defined?", "what is the next free ADR number?", "has anyone shipped an ed25519-pin fix recently?"), each with known-correct source documents/rows.
+
+**Done when:** a checked-in eval file (questions + expected sources) + a tiny runner that scores recall@5 against any search endpoint. Spike: output reshapes V4.R24's go/no-go bar.`,
+    kind: 'spike', discipline: 'engineering', priority: 5,
+    est_minutes: 120, manual_degree: 3, credits_reward: 20,
+    touches: ['tests/', 'docs/'], deps: [],
+  },
+  {
+    ref: 'R02',
+    title: 'ADR: GDS retrieval layer — Postgres-resident, lexical-first, rank-scoped, fail-open',
+    description: `**Criterion:** one-call-recall.
+
+Lock the architecture from the 2026-06-12 research report: everything in the existing Postgres (no new services); lexical FTS first, embeddings only after a measured gap (V4.R24 gate); per-chunk min_rank + builder scoping so the index can never become a read-side bypass of ADR 0016/0043; semantic layer fails open to lexical-only.
+
+**Done when:** ADR merged covering schema, chunking (heading-aware + breadcrumbs), ingestion paths (deploy hook + nightly timer), query fusion (RRF), scoping rules, and the rejected alternatives (ParadeDB, standalone vector DBs, local models, rerankers).`,
+    kind: 'decision', discipline: 'engineering', priority: 5,
+    est_minutes: 180, manual_degree: 3, credits_reward: 40,
+    touches: ['docs/adr/'], deps: [],
+  },
+  {
+    ref: 'R03',
+    title: 'Idempotent ship re-drive — a failed/drifted ship leaves the task one command from resume',
+    description: `**Criterion:** failed-ship-never-strands. Promotes ideas 148 + 175.
+
+The June corpus shows claim_consumed_on_fail firing in 11/11 deep-dived sessions: drift fail → claim consumed → task stuck at completed → can't reclaim → manual confirm or reset-to-ready dance. Fix the state machine: a failed verify/grade/merge leg must NOT consume the claim; ship.js re-run on a completed task with the same builder resumes from the failed leg.
+
+**Done when:** forced-drift and forced-grade-fail both leave the task resumable via one ship.js invocation, with the claim intact; covered by hermetic tests on the ship state transitions.`,
+    kind: 'feature', discipline: 'engineering', priority: 5,
+    est_minutes: 300, manual_degree: 2, credits_reward: 60,
+    touches: ['scripts/gds/ship.js', 'src/bongos/routes/tasks.js', 'src/bongos/db.js'],
+    deps: [], promoted_idea_id: 148, also_promote_ideas: [175],
+  },
+  {
+    ref: 'R04',
+    title: 'Claim-time context pack (static): worktree root, held claims, API enums, OS facts',
+    description: `**Criterion:** sessions-start-knowing.
+
+On /builder-claim (and surfaced by /builder-start), print a compact context pack: the session's worktree absolute root (kills the 84-hit wrong-path class), the claims this builder currently holds (kills claim-ownership mix-ups), the GDS API path prefix + task-creation enums (kind list, priority 1-5 — kills the BAD_KIND/P2 guess-and-retry class), and OS toolchain facts (no \`timeout\` on macOS, psql path).
+
+**Done when:** claim.js emits the pack; content is generated from the live route validators (not hand-maintained); recall hits arrive later via V4.R18.`,
+    kind: 'feature', discipline: 'engineering', priority: 5,
+    est_minutes: 240, manual_degree: 2, credits_reward: 40,
+    touches: ['scripts/gds/claim.js', 'scripts/gds/start.js', 'scripts/gds/cli-lib.js'],
+    deps: [],
+  },
+  {
+    ref: 'R05',
+    title: 'Grader bypass ordering — a 0/10 worker-spawn false-FAIL must never reach the builder',
+    description: `**Criterion:** failed-ship-never-strands.
+
+Session 1659 (2026-06-12) shows a "FAIL · blended 0/10 · low confidence" verdict reaching the builder while GRADER_BYPASS_ENABLED=1 — the bypass check appears to sit downstream of the failing subagent spawn. Reorder so bypass (and any future outage detection) short-circuits BEFORE worker spawn; a worker-infra failure must surface as "grader unavailable", never as a 0/10 grade.
+
+**Done when:** bypass short-circuits pre-spawn; worker-spawn failure yields a distinct non-grade outcome; regression test.`,
+    kind: 'bug', discipline: 'engineering', priority: 5,
+    est_minutes: 90, manual_degree: 2, credits_reward: 20,
+    touches: ['src/bongos/grader.js'], deps: [],
+  },
+  {
+    ref: 'R06',
+    title: 'Migration: search_chunks table + tsvector + pg_trgm (lexical substrate)',
+    description: `**Criterion:** one-call-recall. Cross-ref: design in V4.R02.
+
+search_chunks(id, source_type, source_ref, builder_id NULL, min_rank, track, breadcrumb, content, content_hash, tsv GENERATED, metadata jsonb, indexed_at) + GIN on tsv + CREATE EXTENSION pg_trgm. No pgvector yet (that is V4.R32, gated).
+
+**Done when:** migration applies cleanly on prod + fresh restore; columns support the scoping rules in the ADR.`,
+    kind: 'infra', discipline: 'engineering', priority: 5,
+    est_minutes: 120, manual_degree: 2, credits_reward: 30,
+    touches: ['migrations/'], deps: ['R02'],
+  },
+  {
+    ref: 'R07',
+    title: 'Heading-aware markdown chunker (breadcrumbs, front-matter handling)',
+    description: `**Criterion:** one-call-recall. Cross-ref: V4.R02 design.
+
+Pure module: split md by heading structure into ~300–500-token chunks, prefix each with its breadcrumb path (file > H1 > H2), strip YAML front-matter from the body but keep it as metadata. No fs/network in the core (hermetic tests per the tests/boxes.mjs pattern).
+
+**Done when:** chunker handles CLAUDE.md, ADRs, recipes, session logs; hermetic test suite green.`,
+    kind: 'feature', discipline: 'engineering', priority: 5,
+    est_minutes: 180, manual_degree: 2, credits_reward: 30,
+    touches: ['src/bongos/search-chunker.js', 'tests/'], deps: ['R02'],
+  },
+  {
+    ref: 'R08',
+    title: 'Grade-record reuse on re-ship — never re-run a panel that already passed this diff',
+    description: `**Criterion:** failed-ship-never-strands. Cross-ref: V4.R03 (re-drive).
+
+When a ship re-drive (V4.R03) reaches the grade leg and a stored PASS grade exists for the same diff hash, reuse it instead of burning another panel run (session 1660 did this by hand). Store/compare a diff fingerprint with the grade record.
+
+**Done when:** re-ship with unchanged diff skips the panel and cites the stored grade; changed diff re-grades; tests cover both.`,
+    kind: 'feature', discipline: 'engineering', priority: 5,
+    est_minutes: 240, manual_degree: 2, credits_reward: 40,
+    touches: ['src/bongos/grader.js', 'scripts/gds/ship.js', 'src/bongos/db.js'],
+    deps: ['R03'],
+  },
+
+  // =========================================================================
+  // TIER 3/4 — Build-out (P4)
+  // =========================================================================
+  {
+    ref: 'R09',
+    title: 'Repo-markdown ingester + deploy-hook step (content-hash deltas)',
+    description: `**Criterion:** one-call-recall. Cross-refs: V4.R06 (table), V4.R07 (chunker).
+
+Index all repo md into search_chunks, delta-driven by content hash. Hook: a step in the droplet deploy flow (~/deploy.sh, mirrored in deploy-prod.yml), where new md content + server env are both on disk. Tag chunks from permission-path-protected surfaces (src/bongos/permission-path-check.js globs) with min_rank=metic.
+
+**Done when:** deploy reindexes only changed files; full corpus (~196 files) indexed; protected-surface tagging verified.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 240, manual_degree: 2, credits_reward: 40,
+    touches: ['scripts/gds/search-ingest-md.js', 'docs/recipes/'], deps: ['R06', 'R07'],
+  },
+  {
+    ref: 'R10',
+    title: 'DB-prose ingester + nightly timer — incl. recent commits + shipped-task summaries with recency boost',
+    description: `**Criterion:** one-call-recall. Cross-refs: V4.R06, V4.R07.
+
+Nightly sweep (systemd timer modeled on gds-db-backup) indexing: tasks.description/value_summary, session_logs.notes_md, learnings.body_md, idea_inbox.body_md, task_grades.notes_md, done_when_criteria, blockers — plus recent git commit subjects and freshly-shipped task summaries so "has a parallel session already fixed this?" is answerable (the romantic-gauss redundant-work case). Recency boost in metadata. Session-derived text passes through the ship.js scrubSecrets patterns. builder_memory rows tagged builder_id (owner-scoped); vulnerability_reports excluded.
+
+**Done when:** timer live on the droplet; all listed surfaces indexed; scrubbing + scoping tags verified.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 300, manual_degree: 2, credits_reward: 50,
+    touches: ['scripts/gds/search-ingest-db.js', 'infra/'], deps: ['R06', 'R07'],
+  },
+  {
+    ref: 'R11',
+    title: 'POST /api/gds/search — lexical hybrid (FTS + trgm) fused with RRF',
+    description: `**Criterion:** one-call-recall. Cross-ref: V4.R06.
+
+requireBuilder-gated search endpoint: websearch_to_tsquery leg + pg_trgm leg fused with Reciprocal Rank Fusion in one SQL CTE; k, filters (source_type, track), source citations in results. Per-builder read limiter (memory.js pattern — the global limiter doesn't cap GETs). Add to smoke-gds.js CHECKS (401 unauth).
+
+**Done when:** endpoint live + rate-limited + in smoke; passes the route-rank pre-pass.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 240, manual_degree: 2, credits_reward: 40,
+    touches: ['src/bongos/routes/search.js', 'src/bongos/routes.js', 'scripts/gds/smoke-gds.js'],
+    deps: ['R06'],
+  },
+  {
+    ref: 'R12',
+    title: 'Search rank/owner scoping enforcement + isolation test suite',
+    description: `**Criterion:** one-call-recall (the trust-boundary task). Cross-ref: V4.R11.
+
+Server-side result filtering: min_rank chunks (permission-core content) never served below Metic (ADR 0016/0043 read posture); builder_memory chunks only to their owner; no public/unauthenticated surface. Dedicated isolation suite in the style of tests/memory_isolation.mjs.
+
+**Done when:** isolation suite green, incl. xenos-vs-protected-chunk and cross-builder-memory cases; route-rank pre-pass clean.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 240, manual_degree: 2, credits_reward: 50,
+    touches: ['src/bongos/routes/search.js', 'tests/search_isolation.mjs'], deps: ['R11'],
+  },
+  {
+    ref: 'R13',
+    title: 'Worktree guard — block or loudly warn on writes/git ops crossing worktree ↔ main checkout',
+    description: `**Criterion:** machine-room-ergonomics.
+
+The #1 tool-error class (84 hits): edits/git ops landing in the wrong tree. Add a guard (hook or wrapper) that detects when a write or git command targets a different tree than the session's claimed worktree and refuses/warns with the correct root path.
+
+**Done when:** the wrong-tree write scenario from sessions 196/415/1507 is caught at the moment it happens; opt-out documented for legitimate cross-tree ops.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 180, manual_degree: 2, credits_reward: 30,
+    touches: ['.claude/hooks/', 'scripts/gds/hooks/'], deps: [],
+  },
+  {
+    ref: 'R14',
+    title: 'JSON 404 + hint on API-shaped paths; promote-hint on TASK_NOT_READY 409',
+    description: `**Criterion:** machine-room-ergonomics.
+
+Two self-describing-error fixes: (1) requests to API-shaped paths (/tasks, /gds/*, /api/*) that miss the /api/gds mount get a JSON 404 with a hint instead of the HTML island page (which today causes a secondary JSON.parse fault); (2) the TASK_NOT_READY 409 names the promote endpoint.
+
+**Done when:** both error shapes verified by tests; smoke unaffected.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 90, manual_degree: 2, credits_reward: 15,
+    touches: ['server.js', 'src/bongos/routes/tasks.js'], deps: [],
+  },
+  {
+    ref: 'R15',
+    title: 'Bug: dep-ship auto-promotion flips a BLOCKED task to ready',
+    description: `**Criterion:** failed-ship-never-strands. Promotes idea 226.
+
+The dependency auto-promotion trigger promotes backlog tasks when their last dep ships — but it also flips tasks at status=blocked, silently un-blocking work that is waiting on a blocker. Promotion must only apply to backlog.
+
+**Done when:** trigger guarded; regression test on the blocked case.`,
+    kind: 'bug', discipline: 'engineering', priority: 4,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['migrations/', 'src/bongos/db.js'], deps: [], promoted_idea_id: 226,
+  },
+  {
+    ref: 'R16',
+    title: 'Bug: grader_panel.mjs — 2 failing assertions on a clean main',
+    description: `**Criterion:** failed-ship-never-strands (quality-gate brain must be green). Promotes idea 225.
+
+tests/grader_panel.mjs fails 2/7 on a clean baseline (noted in the 2026-06-08 security-audit session log). Diagnose: stale test vs real grader regression; fix whichever it is.
+
+**Done when:** suite green on main; root cause captured as a learning.`,
+    kind: 'bug', discipline: 'engineering', priority: 4,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['tests/grader_panel.mjs', 'src/bongos/grader.js'], deps: [], promoted_idea_id: 225,
+  },
+  {
+    ref: 'R17',
+    title: '/recall skill — one-call project-knowledge search from any session',
+    description: `**Criterion:** one-call-recall. Cross-refs: V4.R11, V4.R12.
+
+New slash skill wrapping POST /api/gds/search via the cross-platform scripts/gds/api.js: query in, cited sources out, with guidance to follow up by reading the real source. Works identically on laptop, dev box, and browser-only sessions.
+
+**Done when:** skill shipped + documented in CLAUDE.md quick-nav; round-trip verified against prod.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['.claude/skills/recall/', 'CLAUDE.md'], deps: ['R11', 'R12'],
+  },
+  {
+    ref: 'R18',
+    title: 'Claim-time recall hits — top-5 relevant sources printed on /builder-claim',
+    description: `**Criterion:** sessions-start-knowing. Cross-refs: V4.R04 (pack), V4.R11 (endpoint).
+
+Extend the claim-time context pack: query the search endpoint with the claimed task's title + description and print the top-5 hits (ADRs, recipes, session logs, learnings), merged with the existing learnings-by-touches surface.
+
+**Done when:** every claim prints relevant sources; degrades silently when search is unavailable.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['scripts/gds/claim.js'], deps: ['R04', 'R11'],
+  },
+  {
+    ref: 'R19',
+    title: 'Stale-claim auto-expiry/flagging — worktree-less, commit-less claims stop blocking work',
+    description: `**Criterion:** failed-ship-never-strands.
+
+Parallel sessions leave zombie claims (no worktree, no branch, no commits) that block touches-overlapping work and burn reconciliation turns (session 1407). Flag or auto-expire claims with no activity signal after a threshold; surface staleness in /builder-start.
+
+**Done when:** stale claims visibly flagged + reclaimable; expiry policy documented; idle legitimate claims unharmed.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 180, manual_degree: 3, credits_reward: 30,
+    touches: ['src/bongos/db.js', 'src/bongos/routes/claims.js', 'scripts/gds/start.js'], deps: [],
+  },
+  {
+    ref: 'R20',
+    title: 'Laptop smoke preflight — skip or stub server-dependent checks when :3000 is not listening',
+    description: `**Criterion:** machine-room-ergonomics.
+
+Smoke/test legs assume a local game server; laptop sessions hit connection-refused floods (sessions 328, 1113) or missing-module faults. Preflight: detect no listener on 127.0.0.1:3000 and skip server-dependent checks with a clear message (or boot the existing stub).
+
+**Done when:** ship.js smoke leg on a laptop with no server passes with an explicit SKIPPED notice; droplet behavior unchanged.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['scripts/gds/ship.js', 'scripts/gds/smoke-gds.js'], deps: [],
+  },
+  {
+    ref: 'R21',
+    title: 'Digest-time harness-retry sub-count — stop read_before_write/blocked-sleep inflating tool_error_count',
+    description: `**Criterion:** bfg-measures-what-hurts.
+
+Harness guardrail messages (File has not been read yet; Blocked: sleep) are counted as tool errors in session digests, inflating tool_error_count and skewing order=tool_errors (top-2 sessions rank on benign retries). Classify them into a separate harness_retry sub-count at digest build time.
+
+**Done when:** new digests carry the split counts; search ordering uses the net figure; documented in ADR 0027 lineage.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['scripts/gds/ship.js'], deps: [],
+  },
+  {
+    ref: 'R22',
+    title: 'session-eval.js: wrong_path_worktree footgun detector + finding template',
+    description: `**Criterion:** bfg-measures-what-hurts.
+
+The detector misses the corpus's #1 error class entirely. Add WRONG_PATH_RE (File does not exist + cwd note / EISDIR / Cannot find module / command not found: psql|timeout), a wrong_path_worktree footgun type, and a FINDING_TEMPLATES entry routing to idea_inbox.
+
+**Done when:** detector fires on sessions 196/415 fixtures; template produces an actionable finding.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 90, manual_degree: 2, credits_reward: 15,
+    touches: ['scripts/gds/session-eval.js', 'tests/'], deps: [],
+  },
+  {
+    ref: 'R23',
+    title: 'computeOverhead: turn/tool-sequence weighting for cache-dominated sessions',
+    description: `**Criterion:** bfg-measures-what-hurts.
+
+Token-weighted overhead scores 150M+-token cache-heavy marathons as 97–99% "work" even when snippets show hours of ceremony — the BFG is blind to exactly the sessions that bleed most. Add a turn-count/tool-sequence-weighted overhead metric alongside the token one.
+
+**Done when:** the 11 June deep-dive sessions re-score with visible ceremony; both metrics reported; verdict logic uses the turn-weighted one.`,
+    kind: 'feature', discipline: 'engineering', priority: 4,
+    est_minutes: 180, manual_degree: 2, credits_reward: 30,
+    touches: ['scripts/gds/session-eval.js', 'tests/'], deps: [],
+  },
+
+  // =========================================================================
+  // TIER 5 — Measure + close the loop (P3)
+  // =========================================================================
+  {
+    ref: 'R24',
+    title: 'Recall eval run + embeddings go/no-go gate',
+    description: `**Criterion:** one-call-recall (the proof + the gate). Cross-refs: V4.R01 (questions), V4.R09–V4.R12 (lexical layer).
+
+Run the golden-question set against the lexical search. If recall@5 ≥ 80%, the semantic arm (V4.R30–R35) may not be worth building — that is a finding, not a failure; Archon decides at this gate. Spike: output reshapes the P2 tier.
+
+**Done when:** scored results published in the task notes + a go/no-go recommendation for V4.R30.`,
+    kind: 'spike', discipline: 'engineering', priority: 3,
+    est_minutes: 180, manual_degree: 3, credits_reward: 30,
+    touches: ['tests/', 'docs/'], deps: ['R01', 'R09', 'R10', 'R11', 'R12'],
+  },
+  {
+    ref: 'R25',
+    title: 'Session-start recall hints (kill-switched, like the memory pull)',
+    description: `**Criterion:** sessions-start-knowing. Cross-refs: V4.R11, V4.R17.
+
+SessionStart hook addition: when a session opens with an active claim, fetch top recall hits for it (kill-switches mirroring OTB_MEMORY_PULL_OFF / OTB_SUBAGENT; silent on failure).
+
+**Done when:** hook live, kill-switched, and adds <2s to session start.`,
+    kind: 'feature', discipline: 'engineering', priority: 3,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['.claude/hooks/session-start.js'], deps: ['R11', 'R17'],
+  },
+  {
+    ref: 'R26',
+    title: 'Next-free-number helper for ADRs + migrations (kills the numbering-collision class)',
+    description: `**Criterion:** sessions-start-knowing. Absorbs the evidence in idea 122 (ADR 0024 collision) + the triple-ADR-0044 collision (sessions 1660/1629) + the stale claim-time migration reservation observed in this planning session (claim 742 was handed 079, already applied on prod).
+
+One deterministic command (and claim-time print) that returns the next free ADR number and the next free migration number, checked against origin/main + applied schema_migrations, not just local files.
+
+**Done when:** helper shipped + wired into the claim-time context pack; reservation logic fixed or superseded.`,
+    kind: 'feature', discipline: 'engineering', priority: 3,
+    est_minutes: 60, manual_degree: 2, credits_reward: 10,
+    touches: ['scripts/gds/next-number.js', 'scripts/gds/claim.js'], deps: [],
+  },
+  {
+    ref: 'R27',
+    title: 'Session-corpus coverage reconciliation — flag shipped-vs-logged delta',
+    description: `**Criterion:** bfg-measures-what-hurts. Promotes idea 178.
+
+Reconcile shipped tasks against uploaded session records (BFG 6D follow-up to tasks 608/617): every ship should have a digest; flag the delta so corpus gaps can't silently bias the evaluator.
+
+**Done when:** nightly check reports coverage %; gaps listed with task ids.`,
+    kind: 'feature', discipline: 'engineering', priority: 3,
+    est_minutes: 180, manual_degree: 2, credits_reward: 30,
+    touches: ['scripts/gds/session-eval.js', '.claude/scheduled-tasks/bfg-session-eval/'],
+    deps: [], promoted_idea_id: 178,
+  },
+
+  // =========================================================================
+  // TIER 6 — Gated semantic arm + observability build-out (P2)
+  // =========================================================================
+  {
+    ref: 'R30',
+    title: 'Embedding provider decision + key provisioning (Archon go-ahead)',
+    description: `**Criterion:** one-call-recall (gated by V4.R24's go/no-go).
+
+If the eval gate says GO: pick the provider (research recommendation: Voyage voyage-4-lite — $0.02/M tokens, 200M free tier, Matryoshka dims; fallback OpenAI text-embedding-3-small), create the account, provision EMBEDDINGS_API_KEY into /etc/amazonprimea/secrets.env per ADR 0022. New external account = Lars's explicit go-ahead per the budget rules.
+
+**Done when:** decision recorded (ADR addendum), key live on the droplet, cost expectations logged.`,
+    kind: 'decision', discipline: 'engineering', priority: 2,
+    est_minutes: 60, manual_degree: 4, credits_reward: 10,
+    touches: ['docs/adr/'], deps: ['R24'],
+  },
+  {
+    ref: 'R31',
+    title: 'EmbeddingsApi client (do-api.js shape: injectable fetch, dry-run default)',
+    description: `**Criterion:** one-call-recall. Cross-ref: V4.R30.
+
+Thin client mirroring src/bongos/do-api.js: token from env (EMBEDDINGS_API_KEY, ~/.config/otb fallback), injectable fetchImpl, dry-run default, typed error class, batching.
+
+**Done when:** hermetic tests green; no network in tests.`,
+    kind: 'feature', discipline: 'engineering', priority: 2,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['src/bongos/embeddings-api.js', 'tests/'], deps: ['R30'],
+  },
+  {
+    ref: 'R32',
+    title: 'pgvector migration (halfvec + HNSW) + apt-install operator recipe',
+    description: `**Criterion:** one-call-recall. Cross-refs: V4.R06, V4.R30.
+
+ALTER search_chunks ADD embedding halfvec(1024); HNSW index built with bounded maintenance_work_mem (~256MB — never multi-GB on the 2GB box). CREATE EXTENSION vector requires the postgresql-16-pgvector apt package — an out-of-band operator step migrations cannot do; write the recipe.
+
+**Done when:** migration + recipe shipped; extension live on prod; index built off-peak.`,
+    kind: 'infra', discipline: 'engineering', priority: 2,
+    est_minutes: 120, manual_degree: 3, credits_reward: 25,
+    touches: ['migrations/', 'docs/recipes/'], deps: ['R06', 'R30'],
+  },
+  {
+    ref: 'R33',
+    title: 'Embedding backfill + incremental embedding in both ingesters',
+    description: `**Criterion:** one-call-recall. Cross-refs: V4.R31, V4.R32.
+
+Batch-embed the existing corpus (one-time, ~free under the provider free tier) and wire incremental embedding (hash-gated) into the md + DB ingesters.
+
+**Done when:** all chunks embedded; deltas embed on the next ingest run; cost logged via /builder-cost.`,
+    kind: 'feature', discipline: 'engineering', priority: 2,
+    est_minutes: 180, manual_degree: 2, credits_reward: 30,
+    touches: ['scripts/gds/search-ingest-md.js', 'scripts/gds/search-ingest-db.js'],
+    deps: ['R31', 'R32'],
+  },
+  {
+    ref: 'R34',
+    title: 'Hybrid search: vector leg in the RRF fusion + lexical fail-open',
+    description: `**Criterion:** one-call-recall. Cross-refs: V4.R11, V4.R32, V4.R33.
+
+Add the vector leg (embed the query, cosine top-k) to the RRF fusion. HARD requirement: if the embeddings API is down/unconfigured, the endpoint serves lexical-only — semantic is an enhancement, never a single point of failure.
+
+**Done when:** hybrid live; fail-open verified by pulling the key in a test environment.`,
+    kind: 'feature', discipline: 'engineering', priority: 2,
+    est_minutes: 240, manual_degree: 2, credits_reward: 40,
+    touches: ['src/bongos/routes/search.js'], deps: ['R11', 'R32', 'R33'],
+  },
+  {
+    ref: 'R35',
+    title: 'Eval re-run: hybrid vs lexical on the golden set',
+    description: `**Criterion:** one-call-recall. Cross-refs: V4.R24, V4.R34.
+
+Re-score the golden questions against hybrid; publish the delta vs the lexical baseline. If hybrid doesn't beat lexical meaningfully, recommend reverting to lexical-only (keep the layer honest).
+
+**Done when:** comparison published in task notes; recommendation recorded.`,
+    kind: 'feature', discipline: 'engineering', priority: 2,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['tests/', 'docs/'], deps: ['R24', 'R34'],
+  },
+  {
+    ref: 'R36',
+    title: 'Search-index staleness monitoring + /watch tile',
+    description: `**Criterion:** one-pane-of-glass. Cross-refs: V4.R09, V4.R10 (the task-608 staleness pattern).
+
+Visible signals when the index goes stale (last ingest age, chunk counts by source, ingest failures) + a tile on the Archon /watch page.
+
+**Done when:** a deliberately-stalled ingest is visible within a day; tile live.`,
+    kind: 'feature', discipline: 'engineering', priority: 2,
+    est_minutes: 180, manual_degree: 2, credits_reward: 30,
+    touches: ['src/bongos/routes/search.js', 'public-builders/'], deps: ['R09', 'R10'],
+  },
+  {
+    ref: 'R37',
+    title: 'Nightly-dump size threshold review (search_chunks fattens the backup)',
+    description: `**Criterion:** one-pane-of-glass. Cross-ref: V4.R06.
+
+The chunks (+ later embeddings) table inflates the nightly pg_dump from <1MB to tens of MB. Revisit the dump-size expectations/alerts in the backup scripts + recipe so growth is intentional, not silent.
+
+**Done when:** thresholds updated; offsite-copy budget rechecked (ADR 0025).`,
+    kind: 'infra', discipline: 'engineering', priority: 2,
+    est_minutes: 60, manual_degree: 2, credits_reward: 10,
+    touches: ['scripts/gds/db-backup-nightly.sh', 'docs/recipes/gds-db-backup.md'], deps: ['R06'],
+  },
+
+  // =========================================================================
+  // TIER 7 — Proofs + lockdown (P1)
+  // =========================================================================
+  {
+    ref: 'R38',
+    title: 'Proof: forced-drift + forced-grade-fail walkthrough resumes in ≤1 command',
+    description: `**Criterion:** failed-ship-never-strands (the live proof). Cross-refs: V4.R03, V4.R08.
+
+On a staged throwaway task: force a touches[] drift fail, then a grade fail, and demonstrate both resume via a single ship.js invocation with the claim intact and the stored grade reused where valid. Capture the transcript in the task notes.
+
+**Done when:** walkthrough documented; claim_consumed_on_fail does not fire.`,
+    kind: 'feature', discipline: 'engineering', priority: 1,
+    est_minutes: 120, manual_degree: 3, credits_reward: 25,
+    touches: ['docs/'], deps: ['R03', 'R08'],
+  },
+  {
+    ref: 'R39',
+    title: 'Proof: re-run the 2026-06-12 friction analysis — rediscovery, wrong-path, claim-consumed rates drop',
+    description: `**Criterion:** bfg-measures-what-hurts (proving sessions-start-knowing + machine-room-ergonomics too).
+
+Re-run the BFG corpus friction analysis on post-V4 sessions and compare against the 2026-06-12 baseline (1,349 records; taxonomies in the V4.R00 session log): wrong-path hits, guess-and-retry API errors, claim_consumed_on_fail, rediscovery turns. Publish the before/after.
+
+**Done when:** comparison published; criteria 2–4 judged satisfied or follow-ups filed.`,
+    kind: 'feature', discipline: 'engineering', priority: 1,
+    est_minutes: 240, manual_degree: 3, credits_reward: 40,
+    touches: ['docs/'], deps: ['R04', 'R03', 'R13', 'R14', 'R21', 'R22', 'R23'],
+  },
+  {
+    ref: 'R40',
+    title: 'End-of-version docs/index honesty sweep (CLAUDE.md, file-map, architecture, primer)',
+    description: `**Criterion:** cross-cutting lockdown (the task-552 pattern).
+
+CLAUDE.md quick-nav row for /recall + the search layer, file-map entries for the new modules, architecture.md section, builder primer mention, ADR README index. Keep the diagrams honest if any rank/status surface changed.
+
+**Done when:** all indexes reference the shipped V4 surfaces; no stale claims about retrieval.`,
+    kind: 'infra', discipline: 'engineering', priority: 1,
+    est_minutes: 120, manual_degree: 2, credits_reward: 20,
+    touches: ['CLAUDE.md', 'docs/file-map.md', 'docs/architecture.md', 'docs/adr/README.md'],
+    deps: ['R17', 'R38', 'R39'],
+  },
+];
+
+// ---------------------------------------------------------------------------
+// Main: post each task; PATCH idea_inbox promotions; wire dependencies.
+// ---------------------------------------------------------------------------
+
+async function main() {
+  await requireSession();
+  const api = await cliClient();
+
+  const list = await api.tasks.getTasks({ query: { version: 'GDS-V4' } });
+  const bySourceRef = {};
+  for (const t of list.data?.tasks || []) {
+    if (t.source_ref) bySourceRef[t.source_ref] = t;
+  }
+
+  const idByRef = {};
+  let created = 0, skipped = 0, errors = 0, promoted = 0;
+
+  for (const def of TASKS) {
+    const source_ref = `gds-v4-${def.ref.toLowerCase()}`;
+    if (bySourceRef[source_ref]) {
+      idByRef[def.ref] = Number(bySourceRef[source_ref].id);
+      console.log(`SKIP  ${def.ref} (already exists: id=${bySourceRef[source_ref].id})`);
+      skipped++;
+      continue;
+    }
+    const body = {
+      version_id: 'GDS-V4',
+      title: `V4.${def.ref} — ${def.title}`,
+      description: def.description,
+      kind: def.kind,
+      discipline: def.discipline,
+      status: 'backlog',
+      priority: def.priority,
+      est_minutes: def.est_minutes,
+      manual_degree: def.manual_degree,
+      credits_reward: def.credits_reward,
+      touches: def.touches || [],
+      source: 'planning-session',
+      source_ref,
+    };
+    if (DRY) {
+      console.log(`DRY   ${def.ref}  P${def.priority}  ${def.kind}  deps=[${(def.deps || []).join(',')}]  ${def.title.slice(0, 60)}`);
+      continue;
+    }
+    const res = await api.tasks.postTasks({ body });
+    if (res.status !== 201) {
+      console.error(`FAIL  ${def.ref}: status=${res.status} ${JSON.stringify(res.data).slice(0, 200)}`);
+      errors++;
+      continue;
+    }
+    const newId = Number(res.data.task.id);
+    idByRef[def.ref] = newId;
+    console.log(`OK    ${def.ref} → task id=${newId}`);
+    created++;
+
+    const ideaIds = [
+      ...(def.promoted_idea_id ? [def.promoted_idea_id] : []),
+      ...(def.also_promote_ideas || []),
+    ];
+    for (const ideaId of ideaIds) {
+      const patch = await api.inbox.patchInboxId({ id: ideaId, body: {
+        status: 'promoted',
+        promoted_to_task_id: newId,
+      } });
+      if (patch.status === 200) {
+        console.log(`      promoted idea_inbox #${ideaId} → task ${newId}`);
+        promoted++;
+      } else {
+        console.error(`      FAIL promote idea #${ideaId}: ${patch.status} ${JSON.stringify(patch.data).slice(0, 200)}`);
+      }
+    }
+  }
+
+  // Dependency wiring (skipped entirely in dry-run).
+  let depsWired = 0, depErrors = 0;
+  if (!DRY) {
+    for (const def of TASKS) {
+      const deps = def.deps || [];
+      if (deps.length === 0) continue;
+      const taskId = idByRef[def.ref];
+      if (!taskId) { console.error(`DEPS  ${def.ref}: no task id resolved, skipping`); depErrors++; continue; }
+      const depIds = deps.map((r) => idByRef[r]).filter((n) => Number.isFinite(n));
+      if (depIds.length !== deps.length) {
+        console.error(`DEPS  ${def.ref}: unresolved refs in [${deps.join(',')}]`);
+        depErrors++;
+        continue;
+      }
+      const res = await api.tasks.postTasksIdDependencies({ id: taskId, body: {
+        depends_on_task_ids: depIds,
+      } });
+      if (res.status === 200) {
+        console.log(`DEPS  ${def.ref} ← [${deps.join(',')}]`);
+        depsWired++;
+      } else {
+        console.error(`DEPS FAIL ${def.ref}: ${res.status} ${JSON.stringify(res.data).slice(0, 200)}`);
+        depErrors++;
+      }
+    }
+  }
+
+  console.log(`\nDONE  created=${created} skipped=${skipped} errors=${errors} ideas_promoted=${promoted} deps_wired=${depsWired} dep_errors=${depErrors}  (manifest: ${TASKS.length})`);
+  process.exit(errors === 0 && depErrors === 0 ? 0 : 1);
+}
+
+main().catch((err) => {
+  console.error('FATAL', err);
+  process.exit(1);
+});