Skip to content

t27b: the greedy blockers order, its replay and FNV-1a come from specs/tri/t27b/blockers.t27 (Closes #7673) - #7679

Merged
gHashTag merged 1 commit into
masterfrom
port/t27b-blockers
Oct 8, 2026
Merged

t27b: the greedy blockers order, its replay and FNV-1a come from specs/tri/t27b/blockers.t27 (Closes #7673)#7679
gHashTag merged 1 commit into
masterfrom
port/t27b-blockers

Conversation

@gHashTag

@gHashTag gHashTag commented Oct 8, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

Closes #7673. Port slice of #6198 (t27b written in t27), under #6063 and #5980.

What moved, and why this part

cli/t27b/src/blockers.rs is mostly I/O. It spawns and times out processes, runs the reference path (t27c test-report), and reads and appends that path's cache file. Its text parsers build Strings. The blocker names themselves (ExprCall(@abs), type u128) come from lower.rs, not from this file.

Its pure decision logic was three functions, and this PR moves all three into specs/tri/t27b/blockers.t27:

  • greedy: the set-cover order of t27b corpus --blockers;
  • replay: an order's cumulative count over another list of files;
  • fnv64: FNV-1a 64, the hash behind every reference-cache key.

In the spec, the files are the rows of a 0/1 matrix over the constructs, numbered in name order. The tie-break order is unchanged: files unlocked, then the progress score (summed in file order, so the f64 sums are bit-identical), then the files needing the construct, then the name.

t27c gen-rust writes gen/rust/tri/t27b/blockers.rs, which is mounted with #[path] in blockers.rs. The hand-written bodies are deleted. The glue names the constructs and lays out the matrix. The public signatures stay, so main.rs, tests/blockers.rs and tests/refcache.rs do not change.

Lines

lines
hand Rust deleted (cli/t27b/src/blockers.rs) 69
hand glue added (cli/t27b/src/blockers.rs, the only foreign file) 26
net hand Rust -43 (719 -> 676)
generated (gen/rust/tri/t27b/blockers.rs) 106
.t27 (specs/tri/t27b/blockers.t27) 410
seal (.trinity/seals/t27b_T27bBlockers.json) 20

The AGENTS.md remainder line goes from 14919 plus 7874 to 14876 plus 7874 (wc -l).

Gate

Both gates were run on the t27c lab with master's gen/c/policy/own_language.c, the same way as lefthook.yml.

  • check_budget on git diff --numstat --no-renames origin/master...HEAD: exit 0. That is 26 added foreign lines in one file.
  • check_all with origin/master:tools/policy/foreign-exceptions.txt: exit 1, cli/t27b/src/blockers.rs: hand-written .rs is forbidden. The file is not on the exception list.

Evidence that nothing changed

  • Oracle. Every order, count and hash in the spec's 15 tests was printed by the hand-written greedy, replay and fnv64 (master 0fbb0a0, t27c Railway lab).

  • t27c test-report. 15 of 15 pass, 0 vacuous.

  • tri mutate spec. 62 mutants:

    • 55 killed, 0 survived;
    • 6 hung, each a dropped loop step that makes its loop endless;
    • 1 unviable: zig rejects replay with an empty body.

  • Seal. t27c seal --save, then --verify: all hashes MATCH. gen_hash_rust equals the committed generated file.

  • Differential (t27c lab, a throwaway test that is not committed). Master's hand-written greedy, replay and fnv64, copied verbatim, against the generated ones, on:

    • 50000 random file lists (295384 greedy steps), with replays over other lists and over reversed orders;
    • 2000 byte strings;
    • the 519 blocked files of the lab's corpus run (167 constructs), plus their 54-file reference-ok subset.

    Every result is identical.

  • cargo test --release -p t27b --test blockers --test refcache on the t27c lab: 17 passed.

t27b lab corpus runs

I ran two private runs of the deployed harness on the t27b Railway lab, at the same base and with the same settings:

  • the harness is /opt/t27b-lab/lab.py, lab_run;
  • each run has its own T27_WORK and T27_SRV under /tmp;
  • T27_JOBS=12 and T27_REFERENCE_JOBS=3;
  • the corpus step is qemu-aarch64 t27b corpus specs --json --runner qemu-aarch64 --timeout-ms 60000 --jobs 12, and the reference is t27c test-report with zig 0.16.0.
master ef26684 head 94db5f4
files 1578 1579
pass 894 895
pass_vacuous 151 151
blocked 511 511
fail 18 18
timeout 4 4
mismatch 0 0
reference pass 1093 1094
reference disagree 0 0
cargo test t27b (aarch64) 132 passed, 0 failed 132 passed, 0 failed
  • File by file. The 1578 common files are identical in every field: verdict, reference verdict, the full blocker list, tests, invariants, asserts, per-test verdicts and reference disagreements.
  • Timeouts. The same 4 fpga testbenches time out on both sides.
  • The one new file. specs/tri/t27b/blockers.t27 passes in t27b itself (15 tests, 86 asserts) and in the reference. It is unlisted in the ledger, as a64.t27 and macho.t27 are, so the ledger is untouched.
  • Public master run. The lab's scheduled run of master ef26684 (https://t27b-lab-production.up.railway.app/runs/ef26684a1419b31a90c564f303d21998e723224a.json) has the same summary as the private master run.

gen-rust gaps

No new gaps. One known gap shaped the design: a read-only slice parameter other than []const u8 becomes an owned Vec. So replay takes its order as []const usize and the glue moves a fresh Vec<usize> in. The other slices are &[u8] (read-only bytes) or &mut [T] (written).

🤖 Generated with Claude Code

Owner approval (owner-approved-foreign)

  • Who and when. The owner (account gHashTag) applied the owner-approved-foreign label to this PR on 2026-10-08 at 04:44:26Z. The owner did it from their own terminal (gh pr edit 7679 --add-label owner-approved-foreign), after being asked in the working session whether to approve the edit to cli/t27b/src/blockers.rs.
  • Why the label is needed. That file is not on tools/policy/foreign-exceptions.txt.
  • What it does not change. The budget is unchanged: 26 added hand-written lines in one file, 69 deleted, and check_budget() exits 0.

@claude
t27b: the greedy blockers order, its replay and FNV-1a come from spec…
94db5f4 
…s/tri/t27b/blockers.t27 (Closes #7673)

Port slice of #6198. cli/t27b/src/blockers.rs is mostly I/O (processes,
the reference runner, its cache file); its pure decision logic was greedy,
replay and fnv64. They are now specs/tri/t27b/blockers.t27 (15 tests, every
expected value printed by the hand-written Rust on master 0fbb0a0), and
gen/rust/tri/t27b/blockers.rs is its t27c gen-rust output, mounted with
#[path]. The hand-written bodies are deleted; the glue names the constructs
in name order and lays the files out as rows of 0/1 for the spec.

blockers.rs: 26 lines added, 69 deleted (719 -> 676). The public
signatures stay, so main.rs and the tests do not change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@gHashTag
gHashTag enabled auto-merge (squash) October 8, 2026 04:38
@github-actions

github-actions Bot commented Oct 8, 2026

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-10-08 04:38:58 UTC

Summary

Status Count
Total Open PRs 50
PRs with Failing Checks 47
PRs with All Checks Green 3
READY 2
FAILING 47
PENDING 0
NO CHECKS YET 0

These columns do not partition: 2 + 47 + 0 + 0 = 49, and there are 50 open PRs. A PR is being counted twice or not at all.

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=39fa3908f9fd != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

@gHashTag gHashTag added the owner-approved-foreign Owner-approved exception to the only-t27 rule: hand-written foreign code allowed in this PR label Oct 8, 2026
@gHashTag
gHashTag merged commit f917561 into master Oct 8, 2026
40 of 42 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

owner-approved-foreign Owner-approved exception to the only-t27 rule: hand-written foreign code allowed in this PR

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

t27b: the greedy blockers order, its replay and FNV-1a come from specs/tri/t27b/blockers.t27 (port slice of #6198)

2 participants

@gHashTag @claude