Phase C2: per-hit tail honesty for search — the fold where a real query's evidence runs out (D2 Step 2)

[samkeen]

What this is

Split out of #201, which shipped the query-level half of D2 (PR #205) and left this half
untouched by design. Step 1 answers does the vault hold evidence for this query at all. This is
the other question: a real query whose list runs out before limit does — the "results 7–10 are
filler" complaint, the per-hit sibling of the same dishonesty.

Why it was not shipped with #201

The evidence exists and is measured; the labels do not. evals/queries.json names the
relevant note for a query — it says nothing about the irrelevance of ranks 5–10, which is exactly
what a tail rule must be judged on. So the provenance is reported and no rule is drawn from it:

• HitProvenance per served row (bm25_rank, vector_rank, distance), reaching a caller through
  Vault::search_evidence → EvidencedResult.
• just eval's search evidence bake-off prints the dense_only reading — served rows the lexical
  half never ranked. At the shipped bar: 0 of 410 served positive rows are dense-only, against 20
  of 50 negative ones. That separation is the reason to think a tail rule is findable; it is not
  evidence that any particular one is right.

What this issue owes

1. A label extension first (process rule 2 applies: its own commit, and the two-direction token
   audit). The shape is per-query per-rank judgements over the served list, not new notes — the
   corpus does not need to grow, the labels need to get deeper. Decide the cheapest honest form:
   full per-rank relevance, or a per-query "last relevant rank" mark.
2. The rule, bake-off style — swept and re-derived per run like Phase B: the discovery disclosure bake-off — a default-view fold for b2 similar (runs #197 Phase 2, on D1 as redrafted) #200's fold and Phase C: search evidence honesty — b2 search answers zero when the vault holds no evidence (implements D2) #201's bar, on
   both corpora, with "no tail fold at all" an admissible winner (it won for discovery in Phase B: the discovery disclosure bake-off — a default-view fold for b2 similar (runs #197 Phase 2, on D1 as redrafted) #200).
3. D1's prefix requirement binds here exactly as in discovery: per-hit eligibility may not punch
   holes in the fused order. The fold is a cut, or it does not ship.
4. Process rule 5's transfer check if the rule carries any distributional constant —
   just calibrate --search is the bench, and it is the one that caught Phase C: search evidence honesty — b2 search answers zero when the vault holds no evidence (implements D2) #201's first rule.

Judged on

Bench	The reading
orthogonal corpus	zero labelled relevant notes below the tail fold — no headroom (D2's tripwire, per-hit form)
dense fixture	a single-domain vault's lists are all real matches; a tail rule that truncates them there is disqualified, the same absolute #200 enforced for discovery
just calibrate --search on real vaults	process rule 5

Out of scope

• The query-level rule — shipped, Phase C: search evidence honesty — b2 search answers zero when the vault holds no evidence (implements D2) #201.
• The surfaces — Phase D: the search evidence verdict reaches the surfaces — CLI, JSON, desktop, and the exit-gate moves (lands #201; #200 shipped no fold) #202 owns every rendering of a disclosure boundary, per Discovery: surface the "buried gem" that whole-note ranking hides #182's rule.

References

• Phase C: search evidence honesty — b2 search answers zero when the vault holds no evidence (implements D2) #201 (the parent; Step 1's engine work, PR GH #201: Search evidence bar — query-level rule for hybrid_search #205), Phase D: the search evidence verdict reaches the surfaces — CLI, JSON, desktop, and the exit-gate moves (lands #201; #200 shipped no fold) #202 (the surfaces), Phase B: the discovery disclosure bake-off — a default-view fold for b2 similar (runs #197 Phase 2, on D1 as redrafted) #200 (the fold bake-off
  whose method this follows, and whose answer was "no rule"), Discovery floor: the centroid z rule suppresses genuinely-related multi-topic notes outright, and its stated calibration window no longer holds #187 (constants in code, measurements
  in the harness).
• invariants.md D2's closing sentence, which names this as deliberately unshipped;
  docs/evals/README.md process rules 2 and 5.