docs(c-01): measure what a deferred title does to FTS rank and ordering

[arkash20]

The row

pm-0918-c-01 (CRITICAL) argues PersonaMem's 85.4% record (8-10 Sep) was measured against a store that had not finished enriching, so it was never a reproducible baseline. Claimed mechanism: migration 034_memories_search_vector_title_weighting puts title into the FTS search_vector at content's weight and widens the trigger to UPDATE OF content, title, so a deferred-enrichment title rebuilds that row's vector after the row is already searchable — by a non-uniform amount, reordering neighbours.

The row's own decisive test needs prod read access. This PR tests the mechanism offline, read-only, at zero cost: no prod, no seed, no LLM call, no write to the shared local databases.

Verdict

claim	verdict
034 puts title in the tsvector at content's weight	true
a title-only UPDATE rebuilds search_vector	true, demonstrated against Postgres
the rebuild is non-uniform across rows	true — 188 distinct ratios among 411 moved rows
non-uniform ⇒ neighbours reorder	true but not automatic (counterexample in the doc)
enrichment is deferred, titles land after the rows exist	true — 27,794 local rows did exactly that
therefore the 8-10 Sep run scored a moving store	untested — no local population reproduces that trajectory

Mechanism: SOUND. Cause of the 85.4%: still open, and this corpus cannot close it.

Numbers

Temp-table probe borrowing the real trigger function, content never touched:

state	row 1	row 2
no title	0.033333	0.025000
echoing title	0.091667 (x2.75)	0.083333 (x3.33)
non-echoing title	0.033333 (x1.00)	—

Tenant dev-9ff0ca, 8,850 rows, 100% titled, 40 two-term queries per set:

	content-derived	title-derived
rows matching before → after	6,042 → 6,233	1,667 → 2,019
rank unchanged	93.2%	83.9%
top-K agreement (K = min(50, pool))	0.970	0.921
top-10 agreement	0.921	0.779
Kendall tau-b	0.943	0.883
top-1 result changed	9 / 33	12 / 24

Reproduced on a disjoint query set (--seed c01-seedB): aggregate agreement stable to ~0.02; the top-1 rate is not (4/26 vs 9/33), so the doc reports 15-27%, not 27%.

Dose-response in title coverage — default at 6.6% coverage shows zero movement of any kind; dev-9ff0ca at 100% shows the churn above. A store draining its enrichment backlog travels that curve with no deploy to blame.

Two corrections to the row

• "Non-uniform therefore reorders" is not automatic. On default, the title-derived set moved every matched row (592/592, ratios 2.17-3.67) and reordered nothing — tau 1.000, top-1 unchanged on all 6 queries. They moved together, past nobody. Spread is necessary, not sufficient.
• The decisive test is unrunnable as written. It asks for the gap between created_at and when the title was written; memories has no updated_at and no migration ever adds one. The doc gives the audit_log substitute (create-detail records the title at insert), which answers the count and the direction but still not the timestamp — so it cannot say whether titles landed before or after 10 Sep.

Stated limits

No PersonaMem data here (no %amb%/%persona% tenant; newest local row 2026-09-08). Neither measured tenant traverses bulk-write → deferred → fully-enriched: dev-9ff0ca is the destination without the transition (titled at insert), default is an early state that stalled. And the sign is unknown — 034 exists because content-only FTS could not find title-distinguished rows, so a finished store could plausibly score higher.

The search-path bisect the row lists as its fallback should not be cancelled on the strength of this document.

Changes

• docs/fts-title-weighting/c01-title-rank-movement-findings.md — findings
• benchmark/c01_title_fts_rank_movement.py — re-runnable measurement; read-only apart from a rolled-back ON COMMIT DROP temp table

No product code touched. ruff check and ruff format --check clean.

🤖 Generated with Claude Code

[github-actions]

Claude Code Review — skipped: PR author 'arkash20' is not a public member of the 'caura-ai' org

[github-actions]

Claude Code Review — skipped: PR author 'arkash20' is not a public member of the 'caura-ai' org