fix: correct evidence_source on findings that claim datasheet provenance they don't have - #37

Five findings across two analyzers reported an evidence_source that contradicts where the value actually came from. Provenance is what tells a reviewer how much a finding is worth, so a wrong tag is worse than a missing one — it launders an API guess or a package average into datasheet evidence.

Over-claiming (3 sites)

• lifecycle_audit LC-005 (single source) and LC-006 (long lead time) hardcoded evidence_source="datasheet". Both derive purely from distributor API responses — LC-006 reads src_data["lead_time"] and records lead_source="mouser" two lines above the claim that a datasheet was the evidence. No datasheet states a lead time. LC-001..LC-004 in the same file already use "api_lookup".
• analyze_thermal TH-DET and TS-001/002/003/005 set evidence_source="datasheet" when rtheta_ja_source == "package_table". That table is PACKAGE_THERMAL_RESISTANCE — generic per-package Rθ_JA averages selected by a regex over the footprint string. analyze_thermal:394 shows the source is only ever "package_table" or "default", so there was no branch in which "datasheet" was correct and the conditional itself was the bug.

Under-claiming (1 site, the mirror defect)

lifecycle_audit LT-001 mapped any non-empty temperature source to "api_lookup", including "extraction_cache" — which is a real per-MPN datasheet extraction. The same analyzer was over-claiming datasheet provenance in two rules and under-claiming it in a third. Added _temp_evidence_source(): extraction_cache → datasheet, api:* → api_lookup, everything else (e.g. the "mouser_suggestion" tag at :534) → heuristic_rule.

Testing

Validated against kicad-happy-testharness on a 40-repo corpus subset (narrower than the recommended quick_200 = 261 repos), producing 382 thermal outputs.

A/B, main vs this branch, same corpus and same schematic/PCB inputs. Every differing leaf field across all 382 outputs:

48  .assessments[].evidence_source
39  .findings[].evidence_source
35  .trust_summary.by_evidence_source.heuristic_rule
35  .trust_summary.by_evidence_source.datasheet
 7  .audience_summary.designer.top_issues[].evidence_source

Nothing else. No temperature, severity, score or count changed, and zero list-length differences — no detection added or removed. Aggregate shift: {datasheet: 39, geometry: 73, heuristic_rule: 6} → {geometry: 73, heuristic_rule: 45}.

Real-world impact: 39 findings and 48 assessments across 40 repos were claiming datasheet provenance for a footprint-regex package average.

Regression assertions (regression/run_checks.py --type thermal): 9,932 total / 169 passed / 68 failed — identical on main and on this branch, same failure set. Those 68 are pre-existing drift between the seeded assertions and current main, not introduced here.

Re-seeding: none needed. No assertion targets evidence_source as a finding field (0 hits for "field": "evidence_source" across 154,664 assertion files — the evidence_source key in those files is the assertion's own auto_seeded provenance, not the finding's), and baselines/thermal.json records only summary counts with no findings array, verified across all 16,062 thermal baselines. No schema drift: api_lookup and heuristic_rule are both valid in finding_schema.py, and no keys or types change.

Not covered by the harness: lifecycle_audit has no corpus runner (run/ has schematic, pcb, gerbers, thermal, emc, spice, datasheets) — it needs network access and API keys. Its three changes were validated separately on a real 68-component board (24-bit ADC front end, 14 unique MPNs, DigiKey + Mouser + element14 queried): by_evidence_source {datasheet: 7, api_lookup: 6} → {api_lookup: 13}, with trust_level, lifecycle_summary and temperature_summary all unchanged. The new extraction_cache → datasheet branch is unit-tested over all six source tags but not exercised end-to-end, as that board has no datasheets/extracted/ data.

Two related defects found but deliberately left out

These change finding semantics rather than provenance, so they seem like your call rather than mine:

1. lifecycle_audit:731 (LC-ACT) omits confidence and evidence_source entirely — this is what forces trust_level="low" via unknown_confidence — and writes summary "active" for LCSC-only unknown status.
2. lifecycle_audit:744 (LC-005) — total_queried counts only APIs that returned rather than all attempted, status is None counts as active, and LCSC returns no lifecycle status at all, so "only available from X of Y sources checked" is not supported by its own data.

Happy to open either as a follow-up.

Second commit (unrelated, docs only)

skills/bom/references/ordering-and-fabrication.md — the FastAdd example URL ended in &newcart=true. Per DigiKey's own FastAdd guide the default appends to the current cart while newcart=true starts a new empty one, and the cart page warns that creating a new cart deletes the items currently in it. An agent following this reference literally destroys whatever the user had already assembled — and it is not recoverable, since DigiKey exposes no cart read/write API endpoint. Dropped the parameter from the example and documented it as deliberate opt-in.

Also stated the / separator convention for multiple reference designators. The existing paste example already used C1/C2/C5 but never said why, so it read as cosmetic. _write_digikey_order() in bom_manager.py already rewrites commas to /, so generated files were never affected — this only matters for hand-assembled paste blocks.