t27b: per-test reference differential -- mismatch 0 today compares the JIT with an interpreter over the same IR

[gHashTag]

Refs #6434
Refs #6063

Problem

"mismatch 0" on the t27b lab is not a comparison with the reference path.

• In cli/t27b/src/main.rs (the --check block), MISMATCH means the JIT disagrees with eval.rs::Interp. Both read the same IR from lower.rs. If lower.rs mistranslates t27 semantics (wrap vs trap, promotion, float rounding, layout), both sides compute the same wrong thing and agree.
• The only comparison with the reference is per file. blockers.rs::parse_test_report keeps only the tests and FAIL totals of t27c test-report, even though t27c already prints each test's outcome (bootstrap/src/test_report.rs, Report.outcomes).
• steward.t27 is_alarm(fail, fail) is false, so the 7 files where both fail (lab run 03362c1) count as agreement without checking that the same test failed. Example: specs/port/trinity/src/tri/gen_vector2d.t27: t27b fails Vec2_length with assert_eq left 4627730092099942521 vs right 4627730092099895296 (one f64 ulp). The reference says only "1 of 4 tests fail".
• Stop::Fuel and Stop::Depth are also counted as agreement whatever the JIT did.

Change

1. parse_test_report returns the per-test outcome list. The reference cache stores it (new row format; old rows are read as file-level only).
2. t27b corpus --reference compares test by test. A test whose t27b verdict differs from the reference verdict is ref_mismatch, a new verdict counted in summary.ref_mismatch and treated by the doctor as LAB-MISMATCH.
3. steward.t27 is_alarm judges per test when per-test data exists.
4. Interpreter Fuel/Depth stops report unknown, not agreement.

Done when

• The lab JSON carries ref_mismatch per file and in the summary, and the 7 fail/fail files show which tests disagree, if any.
• steward.t27 has tests for the per-test is_alarm.
• New subcommand tri t27b diff <spec> prints both verdict lists side by side.

🤖 Generated with Claude Code

[gHashTag]

First measured reference_disagree on master, from branch t27b-refdiff-20261005T1500 (de50cbd = master 36ad787 plus this change). Run by hand on the Railway t27b lab container at 2026-10-05 15:39-15:47Z: qemu-aarch64 t27b corpus specs --blockers --reference t27c --jobs 12 --timeout-ms 60000. Corpus exit 0.

• reference_disagree 0 files, 0 tests, of 668 files compared test by test. That is 2466 tests per side, and both sides have the same test names.
• jit_interp_mismatch (old mismatch) 0.
• Reference: 722 pass, 25 fail, 509 blocked, 4 timeout. t27b: 380 pass, 281 pass_vacuous, 8 fail, 552 blocked, 37 frontend.
• All 7 files that fail on both sides fail the same single test. gen_vector2d.t27 fails Vec2_length on both, so the file-level agreement was right this time. The others are timing_tb, build_verify, gf16_dot4, check-first-party-doc-language, gen_w392 and search/pattern.
• Not compared: specs/ui/viewport.t27. t27b fails rail_capacity_follows_the_css_budget, but the reference is blocked, so there is no reference verdict to compare it with.

No disagreement was found, so no root-cause issues were filed.

The branch is pushed and has no PR. The owner's rule from 2026-10-05 ~15:45Z forbids hand-written non-t27 code in new PRs. The branch contains the following:

• The steward's decision, in t27: is_alarm_tests and lanes_stop in specs/tri/t27b/steward.t27. gen/c/tri/t27b/steward.c was regenerated, and tri t27b gen-check on the t27c lab says SAME 904af7fc1f22.
• Hand-written glue in Rust (cli/t27b/src/{blockers,main}.rs and tests) and Python (lab.py, t27b.py, t27b_rules.py, CI tests).

Refs #6063