corpus: re-bless the api columns against engine main (gitgalaxy#2729)

[squid-protocol]

Closes #56.
Corpus half of squid-protocol/gitgalaxy#2729, and the paired re-bless owed by
squid-protocol/gitgalaxy#2743 (rosetta:rebless-owed).

100 manifest cells across 27 languages, re-blessed mechanically from one full-precision scan at
engine main 19262fb0 — no hand-edited numbers.

Two causes, both already merged upstream

1. api_orphan_credit 3 → 0 — gitgalaxy#2731, shipped as #2734

25 languages in a/b/c, plus makefile 4 → 3 in c.mk.

api_orphan_credit is arch_api − raw_arch_api: what galaxyscope's Contextual Baseline Fix adds
on top of the api rule's own count when an imported file has uncalled functions. It used to add
all of them, so a function that is both declared public and uncalled counted twice —
data/go/a.go recorded function_count 3, raw_arch_api 3, raw_state_slop_orphans 3 and
arch_api 6. The fix credits only orphans − api_declared_orphans. In this corpus every probe
is declared public by its own language's idiom (that is exactly what #2729's wave planted) and
uncalled in its own file, so the overlap is total and the credit is correctly 0.

New ledger entry api-declared-orphans-are-not-new-surface. The entry that predicted this —
api-double-count-inflates-scored-api, whose verdict said it "resolves when #2731 credits only the
orphans the rule did not already count" — flips still_reproduces to false, with the verifying
numbers: go/a.go arch_api 3 (was 6), risk_api_exposure 3.022 (was 5.939),
risk_documentation 40.613 (was 79.155).

The cost, recorded rather than worked around: the credit was the gate's only assertion that the
main→a→b→c chain resolved — it is how gitgalaxy#2543 and #2540 were caught. 26 of 46 languages now
assert 0 for it. The chain is still asserted by every language's import count and by the 20
languages whose api rule does not reach every probe.

2. api moves — gitgalaxy#2730, shipped as #2743

language	api (main/a/b/c)	why
abap	0 → 4/3/3/3	a FORM subroutine is public by default
ada	0 → 4/3/3/3	a column-0 subprogram is a library-level compilation unit
dart	0 → 4/3/3/3	a top-level function without a leading _
matlab	0 → 4/3/3/3	a function file has no visibility syntax at all — column-0 function is its file-level surface
perl	0 → 4/3/3/3	a sub is package-public, and the only perl form that does not also move a planted state_mutation or import
c	6/9/7/7 → 4/5/3/3	the rule was counting body-local declarations and two-word statements as public surface

c is the "too broad" direction: return NULL; alone was 513 of 8675 matches on the
language-crucible corpus. Its manifest note said "api counts every external-linkage (non-static)
file-scope declaration"; that is now true, where before it also counted every indented one.

api-no-plantable-idiom narrows from seven languages to two, with the reason each survivor
stays:

• objective-c — the rule can now see an @interface method declaration, so the engine half is
  done, but planting an @interface would create a class and SPEC.md rule 2 pins class_start at
  0 as a false-positive tripwire in every language. A corpus constraint, not a rule gap.
• sqlite — unchanged: every spelling its rule accepts (CREATE [TEMP] VIEW, CREATE VIRTUAL TABLE) is a func_start or class_start hit, so there is still no form that does not move a
  planted count.

One plant change: shell

export PROBE_GLOBALS=1 → export -f probe_globals, in all four files.

#2730 made the -f form matchable — the old rule demanded a bare identifier straight after
export, so shell could only ever measure exported variables, and the plant's variable name
merely resembled the probe. api is unchanged at 3 per file; the count was never the point. The
cost is that export -f probe_globals names the function, so under gitgalaxy#2727's
name-recurrence census the probe stops reading as an orphan and api_orphan_credit drops 3 → 0 —
shell joins api-export-names-suppress-orphan-credit, which now records why that trade is worth
making here when it was refused for makefile and m4: it is no longer a distinctive cost, since
25 other languages reach the same place in this same PR, and the alternative is measuring the wrong
construct.

Verification

• verify_language.py — 46/46 PASS
• na_check.py --ci — clean, 0 unreviewed absences
• bias_report.py --gate — exit 0, 0 unexplained out-of-band cells

Ledgered cells fall 194 → 184 and derived 116 → 115: fewer cells need excusing because fewer are
out of band. The headline is raw_arch_api, the column #2729 was filed about — abap and ada
rise 0 → 3.25, c falls 7.25 → 3.75, and risk_api_exposure/risk_documentation come into band
with them. The jcl/risk_api_exposure registry mismatch in the report header is byte-identical to
what is already on main: pre-existing, unchanged, already carried by api-contextual-baseline-fix.

Cross-repo

• fix(core-engine): give the api rule a stated contract and audit all 46 languages against it (#2730) gitgalaxy#2743 (merged) — the engine PR this re-bless is owed to; it carried
  the rosetta:rebless-owed label and its rosetta-audit named these languages before merging.
• arch_api conflates api-rule hits with converted orphans, and the rosetta gate reads the conflated column -- 36 of 44 languages pass api without the rule matching anything gitgalaxy#2729 — closes on this merge; its stated close criterion was
  verify_language.py 46/46 and bias_report.py --gate exit 0, both restored here.
• Nothing to pin: ENGINE_REF/KEYWORD_ROSETTA_REF are retired, and this was blessed against
  engine main, per docs/GATING.md's cross-repo flow.
• On merge, bias-history.yml regenerates the report and closes corpus owes a re-bless against engine main #56; then tools/issue_status.py
  for the languages that moved, per gitgalaxy#2669 F.6.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WfeHRWra1d6Z5uSReFajSq