napkin-math: dropped_signals schema + validator + audit consumption (proposal 141 PR 2)

[neoneye]

Summary

Builds on PR #751 (Fork B advisory audit) by giving the LLM an optional vocabulary to explain absences that would otherwise show up as absent_unexplained in the audit. Three coordinated changes:

1. Extract prompts grow an optional dropped_signals array

Both extract-parameters-from-digest and extract-parameters-from-full get the same schema and rule. Each entry must name a structural reason from a closed enum and reference the current signal it was replaced by, made redundant with, moved to, or capped under.

reason	requires
replaced_by	replacement_id resolves to a current id or output_name
redundant_with	redundant_with_id resolves to a current id or output_name
moved_to_unmodelled_gate	replacement_id resolves to an unmodelled_gates id
cap_pressure	cap_kind names a capped array, AND that array is actually at its cap
out_of_scope	structural rationale only

Hard limit 8 entries; rationale ≤25 words; corpus-agnostic wording (no plan literals).

2. validate_parameters.py ERRORs on malformed dropped_signals

Per the proposal: malformed entries are audit failures, not acceptable explanations. New 19th check dropped_signals_schema enforces the closed enum, resolves references, rejects cap_pressure claims on arrays that aren't at cap, etc. Optional field — absent or empty is the expected default for cleanly-preserved iterations.

3. audit_source_preservation.py reclassifies via strictly validated dropped_signals

New status explained_drop ranks above likely_renamed and absent_unexplained. The audit consumes an entry only when it would also pass the validator — is_audit_consumable_drop mirrors check_dropped_signals_schema so an invalid explanation cannot hide a real regression. Precedence:

preserved_by_id > preserved_by_output_name > preserved_as_formula_dependency
> explained_drop > likely_renamed > absent_unexplained

Specifically, the audit requires for explained_drop:

• origin == "prior_baseline" (source_digest entries are Fork A territory)
• reason in the closed enum
• replaced_by / moved_to_unmodelled_gate → replacement_id resolves
• redundant_with → redundant_with_id resolves
• cap_pressure → cap_kind names a capped array AND that array is actually at cap

Malformed entries fall through to likely_renamed / absent_unexplained instead of being silently reclassified. Double-counting is avoided by ignoring entries whose id is actually preserved in current.

Empirical posture

• 56 unit tests pass (28 validator + 28 audit). 23 new for the schema and consumption; 5 new for strict rejection paths (unresolved replacement_id, wrong origin, unresolved redundant_with_id, unjustified cap_pressure, moved_to_unmodelled_gate pointing at a key_value).
• 9/9 smoke checks pass.
• All 6 v51 parameters.json validate clean with 19 checks (was 18) — the new dropped_signals_schema check is correctly inert on legacy outputs that don't have the optional field.

Discovered limitation (honest surface)

The LLM can only meaningfully emit dropped_signals with origin: "prior_baseline" if the orchestrator passes it the prior parameters.json as additional input. The current extract skill reads only the source digest, so a same-LLM same-session regeneration of v51 would emit zero prior_baseline drops — the LLM has no information about what the prior baseline contained.

This PR therefore ships the schema/validator/audit-consumption infrastructure. The orchestrator/skill wiring that lets the LLM see prior baselines is a candidate for proposal 141 PR 3 (alongside, or before, Fork A). The audit's Fork B comparison itself does not need the LLM to know about the prior — the audit reads both files externally; this PR's job is just to let the audit consume LLM-emitted explanations once they exist.

Out of scope (later proposal 141 PRs)

• Fork A: source-digest regex scan against the current artifact (independent advisory line).
• Orchestrator wiring to pass prior parameters.json to the extract skill.
• Strict mode / CI gating policy.
• source_claim_ids per-entry field (per-claim grounding).

Test plan

• pytest experiments/napkin_math/tests/test_validate_parameters.py experiments/napkin_math/tests/test_audit_source_preservation.py (56 pass; 28 new across both files)
• python3 experiments/napkin_math/tests/run_smoke.py (9/9 pass)
• All 6 v51 plans still validate clean with 19 checks (new schema check is no-op when field absent)
• Audit strictly consumes only valid dropped_signals (mirrors validator checks)
• CI green on latest head

🤖 Generated with Claude Code

[neoneye]

Addressed the strict-consumption blocker (`87a6e7fa`).

Fix: introduced `is_audit_consumable_drop(entry, current_params, current_index)` that mirrors the validator's checks. The audit now consumes a `dropped_signals` entry only when it would also pass validation:

• `origin == "prior_baseline"` (source_digest entries are Fork A territory and not consumed here)
• `reason` in the closed enum
• `id` is non-empty
• `replaced_by` → `replacement_id` resolves to current id or output_name
• `redundant_with` → `redundant_with_id` resolves to current id or output_name
• `moved_to_unmodelled_gate` → `replacement_id` matches an `unmodelled_gates` id
• `cap_pressure` → `cap_kind` names a capped array AND that array is actually at its cap

Malformed entries fall through to `likely_renamed` / `absent_unexplained` instead of being hidden by the invalid explanation.

Five new synthetic tests cover the exact rejection paths you asked for:

• `test_explained_drop_rejects_unresolved_replacement_id` — the literal failing case from your message; the prior signal correctly falls through to `likely_renamed`.
• `test_explained_drop_rejects_source_digest_origin` — wrong origin → no consumption.
• `test_explained_drop_rejects_unresolved_redundant_with_id`
• `test_explained_drop_rejects_unjustified_cap_pressure` — array not at cap → no consumption.
• `test_explained_drop_rejects_moved_to_unmodelled_gate_pointing_at_kv` — pointing at a key_value instead of an unmodelled_gates entry → no consumption.

Also updated the previously-passing positive `cap_pressure` test to fill `key_values` to its cap of 8 so the claim is justified.

Module docstring refresh: `explained_drop` is now listed in the status enumeration, the `dropped_signals` consumption is documented under audit behaviour, and the out-of-scope list correctly excludes `dropped_signals` (which this PR adds).

56 unit tests pass total (28 validator + 28 audit; 5 new for strict consumption). 9/9 smoke checks. CI re-run pending.