Repository navigation
t27b overnight loop 2026-10-05: 100% coverage, broken specs, reference denominator #6434
Closed as not planned
Description
Refs #6063
Overnight improvement loop started by the owner on 2026-10-05 ("merge all t27b PRs yourself, go to 100%, fix every broken spec"). A 15-minute in-session cron runs it; the ledger is ~/.local/state/t27b-queen/ledger.md.
Where it starts (lab run at master 03362c1)
- The reference path (t27c gen-zig + zig test) passes 722 of the 1259 specs in this repo.
- Among those 722: t27b passes 379 with tests checked, and 281 have no tests at all (vacuous pass). 62 are still blocked by t27b. Mismatch 0.
- The other 534 specs fail the reference path itself (509 blocked, 25 fail), so they sit outside the denominator.
- The SPECS catalog at app.t27.ai/game/specs covers 1778 specs from 14 repos. It shows 176 Broken and 232 Warnings. It was built from a vendored WASM at bf7d8ee. 23 of the Broken entries are intentional negative fixtures.
Plan (one lane per row; each lane files its own issue and PR)
- t27b lane: tuples (16 files)
- t27b lane: f32 (7 files)
- t27b lane: tail blockers (undefined, std.*, method calls, assign, alias, anonymous struct, void)
- Broken in gHashTag/t27: re-measure with master t27c; fix the specs; show negative fixtures as expected-fail in the catalog
- Broken in other repos (trinity-fpga, tt-trinity-gamma, turbobaby-user-bot, tt-trinity-euler, trios, tt-trinity-phi, trinity)
- t27c reference denominator: fix(t27c): read a dotted module name and use path whole #6102, then the largest gen-zig failure classes (201 undeclared identifier, gen-zig: brace invariants and tail expressions become 'value ignored' (16 specs do not compile) #6315, t27c gen: open slice x[i..] emitted as x[i]; &[_]T{...} emitted as &.{ _ } #6250, t27c gen (Zig): test bindings shadow module vars, redeclare locals, and inherit param renames #6295)
- tri: a catalog-health command so Broken/Warn counts come from one command, not a hand-written script
- Steward self-heal: a silent ledger is an anomaly (skill rows Q44 and Q45)
- Vacuous passes: 281 reference-passing specs have no test. An owner decision is pending on whether they count.
Weak points found
- The steward was dead from 06:04Z to 14:35Z with no alarm. The claim pid was gone and no ticks ran.
- Q37 (one lower.rs lane at a time) serialised all coverage work. This loop moves new lowering into per-lane submodules instead.
- 42% of the corpus fails the reference itself, so "100% of reference passes" understates the work.
- The SPECS catalog is built by a stale WASM compiler and counts intentional negative fixtures as Broken.
🤖 Generated with Claude Code
Boundary
cli/t27b/src/lower.rs
User Scenarios
- Given the overnight loop runs against master 03362c1, when the tuples lane executes, then t27b passes all 16 tuple-related specs with tests checked
- Given the reference path passes 722 specs, when the loop completes the tuples lane, then the t27b pass count increases from 379 toward the reference denominator without new mismatches
Requirements
- FR-001: The tuples lane MUST lower tuple types and expressions in
cli/t27b/src/lower.rswithout breaking the 379 currently passing specs - FR-002: The 16 tuple-related spec files MUST compile via t27c gen-zig and pass zig test when the lane completes
- FR-003: The lane MUST execute as an independent submodule so Q37 serialisation no longer blocks other lanes
Success Criteria
zig test cli/t27bexits 0 with the 16 tuple specs added to the passing set- The ledger at
~/.local/state/t27b-queen/ledger.mdshows tuples lane completed with 0 new mismatches cli/t27b/src/lower.rscontains the new tuple lowering submodule and no regression in the existing 379 passing specs
Activity
gHashTag commentedon Oct 5, 2026
Tick 15:03Z: lab on master 682564f -- t27b passes 380 of 723 reference-passing files (fail 8, timeout 2, JIT/interp mismatch 0). Merged since last report: #6440 (five catalog-Broken t27 sources parse and generate). Open from this loop: #6439 (tri catalog health). In flight: lanes tuple / f32 / misc-tail, broken-t27, broken-ext, t27c-ref, #6446 gen typecheck gate, #6441 per-test reference diff. Host load is ~217 from other sessions, so no new lane is spawned until one finishes.
gHashTag commentedon Oct 5, 2026
t27b vs other second backends -- 2026-10-05
t27b is the native AArch64 JIT backend of t27. It is measured against the reference path,
t27c gen-zig+zig test.Our numbers (lab, master 682564f)
Comparison
t27b_expectations.json), but the lab image does not run it (#6443)SkipZigTest)pass_vacuous)tri t27b next, rule in steward.t27)tri t27b doctor/watch/ready)Reading the table honestly
gencode generator does not refuse specs that typecheck refuses (t27c gen emits code for specs typecheck refuses (35 misread pairs stay 'green') #6446). This is why the 35 misreads can stay "green".Sources
gHashTag commentedon Oct 5, 2026
Tick 15:16Z: #6439 merged (tri catalog health, 11f5e6a). Anomaly: master Corpus ratchet red since #6440; self-heal PR #6453 open. New agents: bee-feeder (Queen board had 18 bees running and 0 t27b cards) and lab freshness #6443. Lab unchanged at 682564f: t27b passes 380 of the 723 specs the reference passes, mismatch 0.
github-actions commentedon Oct 10, 2026
Re-filed as #8520.
An attempt claimed this issue on 2026-10-07 and left an empty branch. A dispatch row that has spent its retry ceiling is released once and then keeps its issue for good, which is the right rule for a third identical failure and the wrong outcome for an issue nobody ever attempted twice. The new issue carries the same work, the same boundary, and today's brief.
Closing this one so the swarm can take the work again.