fix(search): module-doc chunker + P3 anchor + reranker scale + measurement hygiene (P3 gate GREEN)

[ManSio]

Mechanism (why P3 flipped FALSIFIED -> PASS)

Three stacked changes, each load-bearing (ablation in Red Team):

1. Chunker (src/core/indexing/parser.py): AST chunker dropped module-level
   docstrings index-wide, so the passage the reranker scores highest never
   existed as a chunk. Files with a docstring now get it as chunk 0
   (symbol_type=module_docstring, hierarchy_level=module); no-docstring
   behavior unchanged. Live index: 371 module-doc chunks, all chunk_index=0.
2. Reranker scale (src/providers/reranker/multi_provider.py): llama.cpp raw
   logits (e.g. P3 code chunk -0.99) were filtered against MIN_RERANK_SCORE=0.3
   unnormalized. Sigmoid normalization per the scale contract.
3. P3 pool anchor (src/core/search/engine.py:_anchor_module_head_chunks_async):
   appends each pool file's chunk-0 head pre-rerank (tail, no re-sort; capped by
   _O1_ANCHOR_TOTAL/MAX_RERANKER_INPUT; doc/data files ineligible). P2 precedent.
4. Measurement hygiene (src/core/system_artifacts.py + file_guard.py,
   cherry-picked from fix/index-hygiene-final-gate 496aa04d, clean apply):
   probe/result JSON dumps + experiments/**/results|work excluded from the index
   (they contain verbatim query terms and dominated BM25).

Numbers (FULL reindex + fresh-process gate, rev f1c32d05)

Index (standalone indexer.index_project + physical clear, same call the MCP
intel_trigger_reindex(mode=full) makes — MCP server was down, see caveats):
11026 rows (was 14832), 738 files, IVF_FLAT finalized (was flat,
num_indices=0; now 11026/11026 indexed, COSINE), junk results/work chunks
2152 -> 0, probe/result-JSON chunks 2092 -> 0.

Gate scripts/p3_holdout_gate.py (FTS prebuild + discarded warm-up + per-case
reranker-cache clear), 3/3 fresh-process PASS, bit-identical rows:

P2 rank=1 | P3 rank=1 | R2 rank=1
H1 4 | H2 None | H3 1 | H4 1 | H5 1 | H6 2 | H7 None | H8 2
H9 None | H10 None | H11 1 | H12 1 | N1 clean | DOC clean

vs FALSIFIED baseline (junk index, no chunker/sigmoid):
P3 None, R2 None, H1 None, H3 None, H4 2, H8 1, H11 2, H12 None.
No unexpected boosts, no doc boosts, no reranker-cache voids, no degraded rows
(model=llama.cpp-reranker every query). FTS-timeout flake: 0 occurrences in
3 gate runs + ~10 probe processes (all stderr clean; the timed_out fail-loud
path is present, just never triggered). Unit suites: 34/34 pass
(test_chunker_module_doc 5, test_measurement_hygiene 20, test_p3_module_head_anchor 9).

Red team (GREEN gate attacked, 4 attacks + stability work)

• A1 docstring-vs-code at rank 1 — DEFENDED. P3 Quick question about your repo #1 IS the module-doc chunk
  (artifact_gc.py:0, score 0.77, module_head_anchor=True); prereg clause (i)
  holds live. Anchor-OFF control: P3 Quick question about your repo #1 falls back to a code chunk (0.27) —
  chunker+anchor are load-bearing, not decorative.
• A2 anchor demotion of H-set (prereg ii) — DEFENDED. Ablation ON vs OFF on
  H1/H8: target file-rank unchanged; H8's Quick question about your repo #1 competitor (e17_golden.json,
  anchor=False) enters via BM25/vector tiers, never via the anchor. Anchor exonerated.
• A3 junk-term residual at search level — DEFENDED. ctx_F5S-11_B,
  judged_raw_t5 return hits only from legit indexed manifests
  (INVENTORY.json, MANIFEST.json), 0 hits from excluded results/work paths.
• A4 holdout leakage via indexed writeups — DEFENDED. reranker_p3/README.md
  matches P3 at 0.00016 vs gold 0.77 (~4000x gap); DOC control clean all runs.
• OPEN (medium): H8 file-rank is order-sensitive — 1 in isolation (10/10),
  2 in full-gate order (5/5, bit-identical). Quick question about your repo #1 in gate order is the legitimately
  indexed eval file experiments/bootstrap/e17_golden.json (cypher vocabulary,
  BM25 rank 8 + tiers → rerank 0.17 vs code chunk 0.11). Mechanism unproven
  (prime suspects: RRF pool-boundary interplay + ±5% cross-run reranker score
  wobble observed on identical inputs). No gate impact (H-rows report-only), but
  honest flag. Guard proposals: (a) extend hygiene exclusion to
  experiments/bootstrap/e17_* golden files; (b) gate dumps top-5 per row for
  auditability; (c) 3x repetition of volatile rows in the gate.

Caveats

• MCP server (src.main) was killed mid-task to pick up the synced extension code
  and did not auto-respawn; reindex ran standalone via repo venv with the identical
  indexer.index_project + recreate_table_physical path (log: 738 files, 528.7s).
  Owner action needed: reload Zed to bring the MCP server back on the new code.
• experiments/planted_break/results.json local modification left untouched
  (unrelated dirt, not staged, not in this PR).
• H7/H9/H10 None = threshold cuts, same as baseline (no regression); H1 4 and
  H6 2 reflect genuine topical competition (dedicated FTS5/graph experiment files
  outscore the generic gold chunk).
• Do NOT merge yet — red-team OPEN item + Zed reload pending.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: a8bc1788-351d-453d-8ef9-232bae261000

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[ManSio]

Conflict map vs #62/#63 (for your rebase; owner decision: #58 merges first, I adapt after).

KEEP YOURS (superset, no action needed on my side):

• Measurement-hygiene guard (dirs + probe/result names) supersedes my fix(search): cold FTS build out of 2s budget plus index experiment-output guard #62 is_experiment_output subset; I will drop my impl + test on rebase (already audited: your hygiene tests cover my 7 cases).
• Chunker + P3 module-head anchor + probes/holdout gate: unique, untouched by me.

SWAP ON REBASE (mine wins, #63 already open):

• Your naive _sigmoid (1/(1+exp(-x)), overflows past |x|>~709) vs my numerically stable branching version + 4 tests (incl. P-002 extremes). Please take the stable one; feat(reranker): salvage PR52 scale floor plus holdout calibration #63 will replace it right after your merge.

NO-OP overlaps (disjoint regions, merge cleanly after rebase):

• My FTS-hoist (engine FTS block) vs your chunker/P3-anchor blocks; my top-N floor + calibration (settings/scoring untouched by you); my purge/relang artifacts.

DB note: purge of 772 garbage files already executed against the shared index (verified 0 left); no re-purge needed after either merge.