test(core-engine): the raw/adjusted api split is not a sum (#2729)

[squid-protocol]

Part of #2729 — the last piece of its own "Engine/test surface" half.
Test-only; no production code is touched.

What #2729 asked for here

Engine/test surface: make the split explicit. […] so api asserts what the rule found and
the conversion is asserted separately (as arch_api - raw_arch_api, or not at all).

The gate half landed corpus-side in keyword-rosetta#53 (verify_language.py reads raw_arch_api
as api and pins the conversion separately as api_orphan_credit). #2731/#2734 added the engine
tests for the conversion itself, at both layers that own it — test_detector.py's
api_declared_orphans tests and test_galaxyscope.py's Contextual-Baseline-Fix tests.

One thing was left behind: the recorder test was still documenting the pre-#2731 arithmetic as
an engine invariant.

The stale assertion

test_raw_columns_persist_pre_adjustment_values ended with:

assert row["arch_api"] == row["raw_arch_api"] + row["raw_state_slop_orphans"], (
    "Contextual Baseline Fix invariant broken: adjusted api must equal raw api + raw orphans"
)

Two problems:

1. It never proved that. Both sides are literals fed into the fixture (hit_vector=[5, 0],
   raw_pre_adjustment={"api": 2, "orphaned_logic": 3}). The test passes on its own arithmetic,
   not on the engine's.
2. It is false in general since fix(core-engine): the api rule's own declarations are not new orphan surface (#2731) #2734. The conversion credits
   orphans - api_declared_orphans, so a file whose api rule already declared some of its orphans
   records an adjusted value below that sum — and total overlap (every public function uncalled:
   the library case, and every file in the keyword-rosetta corpus) makes it raw_arch_api exactly.

So the message a reader takes from the only recorder-level statement of the contract is the double
count #2731 removed, written down as an expectation. The fixture happens to be the no-overlap case,
which is why 2 + 3 == 5 holds in it — nothing was catching that.

What replaces it

• The same test now asserts what the raw columns are actually for: the credit is recoverable as
  adjusted - raw, without assuming what it equals. That is the identical subtraction
  keyword-rosetta's gate reads as api_orphan_credit.
• New test_raw_columns_recover_a_partial_conversion_credit covers the shape the sum-based
  assertion excluded by construction — 4 declared, 3 orphans, 2 of them already public → adjusted
  5, credit 1 — and asserts arch_api != raw_arch_api + raw_state_slop_orphans on purpose, so
  a regression back to the pre-arch_api counts a declared-public, uncalled function twice: 3 functions record api 6 #2731 arithmetic fails here rather than only in the corpus.
• The section header now says which layer owns which number, so the next reader does not re-derive
  the conversion from the recorder's columns: galaxyscope.py owns the adjusted value,
  detector.py owns api_declared_orphans, these tests own the round-trip.

Verification

Full suite 7626 passed serially (CI's mode), ruff and mypy clean against their baselines. A
test-only change cannot move the golden masters, the tri-comparison baselines or the tree-sitter
baselines, and none of them moved.

Cross-repo

Companion: squid-protocol/keyword-rosetta#57, the corpus re-bless that restores #2729's stated
close criteria (verify_language.py 46/46, bias_report.py --gate exit 0). Neither PR depends on
the other — this one changes no observable count, so it carries no rosetta:rebless-owed label and
its rosetta-audit classifies the outstanding drift as pre-existing until #57 merges.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WfeHRWra1d6Z5uSReFajSq

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab