docs(96): calibrate the escalation trigger — native-exactness dominates dualpass_flagged

[r-uben]

Summary

Docs only. Closes the open design question left by the Unit A calibration (#104),
which invalidated both load-bearing pieces of the Unit B sketch and said a trigger
"still needs designing from signals that demonstrably fire".

Every candidate is now cross-tabulated against the ground truth of does escalation
actually help this page.

trigger	fires	TP	FP	missed	precision	recall
native-exactness < 100%	13/16	9	4	0	69%	100%
a ground-truth row is missing	13/16	9	4	0	69%	100%
orphan value rows present	4/16	4	0	5	100%	44%
tables declared, none emitted	3/16	3	0	6	100%	33%
dualpass_flagged (current)	16/16	9	7	0	56%	100%

Recommendation: escalate when the emitted table disagrees with its own native layer

A candidate the original sketch never considered turns out to dominate. On a
born-digital page the exactness metric is computable at runtime, model-free, at
zero cost — the native text layer is right there. That is a direct measure of the
defect rather than a structural proxy for it.

It strictly dominates the signal socr flags on today: identical recall (no
missed gains), better precision, three fewer pages fired.

The discrimination is precisely what #95 found missing. Pages 45, 59 and 60 are
100% correct and this trigger stays silent on all three — while dualpass_flagged
fires on them, because it fires on every table page in the document.

Properties that matter for the lane:

• Parameter-free — "disagrees at all", not a tuned threshold. exactness < 100%
  and rows_not_found > 0 select the same 13 pages.
• Model-free and free — no engine call to decide whether to make an engine call.
• Fail-safe in direction — the four false positives cost ~$0.0002 each, and
  escalation was never worse than the incumbent on any page. Over-triggering costs
  money, not correctness.
• Same oracle as the canary, and it degrades coherently: where the native layer
  is a poor oracle the trigger fires and the canary then rejects, so the page is
  left alone at the cost of a wasted call.

Limits, recorded in the note

• Born-digital only. A scanned page has no native oracle, so neither trigger nor
  canary can run.
• One document, no negative controls. 69% is a point estimate on 16 table pages
  from one report; only the ordering against dualpass_flagged is robust.
• It does not detect misalignment as such. It catches the mode-(a) pages because
  the shift also drops rows (parents emptied, orphans trailing). A permutation that
  preserved cell-to-label binding would be invisible to both trigger and canary —
  not a failure mode observed here, but worth knowing.
• Page 53 is a permanent false positive until its native parse is fixed, and is why
  precision is 69% rather than 75%.

Test plan

Not applicable — documentation. Reproducible from the paths in the bake-off note's
Setup table plus benchmark/table_exactness.score_page, which shipped in #104.

Refs #96