Two-column headerless work history is dropped and the probe still reports ok

[s-annam]

Follow-up to #492, split out so the residue that PR #842 does not cover is recorded before #492 closes on merge.

What still breaks

recoverHeaderlessExperience (src/lib/heuristics/sections.ts) returns early on !singleColumn. So a two-column résumé whose work history carries no recognized experience heading is still dropped whole — exactly the #492 symptom, on the layout the recovery deliberately excludes.

The single-column case is fixed by #842 and is not in scope here.

Why it is invisible

#492's AC 5 asked that /probe-experience report a defect (not ok) on a headerless-experience résumé. #842 satisfies that AC by dissolving it for single-column — the fixture now routes an experience section, so ok is the correct verdict. It does not satisfy it for two-column:

• localizeExperience scopes its date-range oracle to the already-routed experience region (src/lib/heuristics/localize/experience.ts:92). With no experience section routed, there is no region, so the oracle never sees the dated role cluster sitting in other.
• EXPERIENCE_DEFECT_CLASSES has only experience-parser-miss and experience-under-segmented. Neither describes "no experience region exists at all".

Net: the probe reports Verify: ok with zero roles parsed, silently. That is worse than the parse miss itself — the lane whose job is to flag the miss reports clean.

Acceptance criteria

1. A new experience-no-section defect class in EXPERIENCE_DEFECT_CLASSES.
2. The oracle runs document-wide (not region-scoped) when no experience section was routed, so a dated title-cased cluster anywhere in the document raises the defect.
3. /probe-experience reports experience-no-section — not ok — on a two-column headerless-experience fixture.
4. A fixture pinning the two-column headerless shape, with a synthetic persona per the tests/fixtures/pdfs/README.md policy.
5. Recovery itself stays out of scope: this issue makes the drop visible, it does not make it parse. Widening recoverHeaderlessExperience past singleColumn is a separate decision — the two-column reader has to distinguish a headerless work-history rail from a sidebar, which the single-column recovery never has to.

Notes

• Non-regression: these résumés were dropped before fix(heuristics): recover dropped roles, titles and skills sections #842 too. This is uncaptured pre-existing behaviour, not new breakage.
• Related: Parser drops entire work history when the experience section has no header #492 (the parent symptom, closed by fix(heuristics): recover dropped roles, titles and skills sections #842 for single-column), [epic] Architecture: max-impact modularization + parsing-accuracy roadmap #646 (the scored-router item that would subsume the region-scoping limitation).

[s-annam]

Cost note, carried over from the session that deferred this.

An earlier pass deliberately did not file this, on the grounds that a new experience-no-section defect class needs a new DerivedSignals key — which rewrites all 60 corpus baselines and bumps the snapshot schema version. That objection is still true and is the main reason this is not a quick fix.

It was filed anyway because Resolves #492 closes the parent on merge of #842, so the two-column residue would otherwise go unrecorded entirely. Filing preserves the finding; it does not argue the cost is small.

Practical consequence for whoever picks this up: batch it with another schema bump rather than paying a 60-baseline rewrite for this alone. AC 4 (a new two-column fixture) is the cheap half and can land first if the defect class is deferred.