t27b: key reference verdicts by the generated code, not the t27c binary

Boundary

specs/tri/t27b/verdict_key.t27 (new), its gen-rust output gen/rust/tri/t27b/verdict_key.rs, its seal, and the call sites in cli/t27b/src/blockers.rs (toolchain_stamp, RefRunner::key, RefRunner::run). No workflow change. The lab's own Python cache (contrib/railway/t27b-lab/lab.py) is not touched here.

User Scenarios

The t27b reference cache keys every row on the bytes of the t27c binary. A pull request that touches bootstrap/ therefore misses every row and re-runs t27c test-report for every spec: run 37911771594 spent 1406 s in the ratchet's corpus step (cache 20539 -> 22249 rows), where the spec-only run 37911494033 spent 432 s (20000 -> 20003 rows). Most compiler changes leave the generated code of most specs unchanged, so their verdicts could be reused.

Requirements

The key is defined in .t27 and generated with gen-rust: FNV-1a 64 over a schema number and five parts, in order: the spec and its use closure; the stdout of t27c gen <spec>; bootstrap/src/test_report.rs (the test runner); zig version; the arch-os from zig env. This is how seal v2 (t27c seal writes the v2 identity, and t27c frontier reports what may be reused (Closes #8104, Closes #8109) #8110) judges a toolchain.
The host is a key part, so a host-dependent verdict (reference: self_attention.t27 passes or fails by host -- its test iterates an undefined W585 scaffold slice (t27b ledger STALE) #8051, self_attention) is never served on another host.
A part that cannot be computed is replaced by the t27c binary hash, so the key is never looser than today.
Never cached: a reference timeout, a t27c that did not exit normally, a test the runner stopped at its timeout, and zig resource faults. A cached run without per-test verdicts stays a miss.
blockers.rs only computes the parts and calls the generated functions, within the 40/80 foreign-line budget (Only-t27 gate: a foreign-line budget the label cannot lift; local hook reads exceptions from origin/master #7371).

Success Criteria

The spec's tests pass under t27c test-report. Mutants (host dropped, timeout check dropped) are killed.
cargo test -p t27b --test refcache passes.
On the lab, a t27c built from t27c seal writes the v2 identity, and t27c frontier reports what may be reused (Closes #8104, Closes #8109) #8110's head reuses the verdicts of every spec whose generated code did not change. The measured hit rate and corpus time go in the pull request.

Refs #8095, #6063

Requirements

FR-001: The key MUST be defined in a .t27 spec and generated with gen-rust as FNV-1a 64 over a schema number and five parts in order: the spec and its use closure; the stdout of t27c gen <spec>; bootstrap/src/test_report.rs (the test runner); zig version; the arch-os from zig env.

FR-002: The host MUST be a key part so that a host-dependent verdict is never served on another host.

FR-003: A part that cannot be computed MUST be replaced by the t27c binary hash so the key is never looser than today.

FR-004: A reference timeout, a t27c that did not exit normally, a test the runner stopped at its timeout, and zig resource faults MUST never be cached. A cached run without per-test verdicts MUST stay a miss.

FR-005: blockers.rs MUST only compute the parts and call the generated functions, within the 40/80 foreign-line budget.

Workflow follow-up for #8208 (header comment) and #8185 (step-summary line). This push token has no workflow scope, so neither PR touches .github/workflows/. Apply by hand once both have merged: git apply this diff on master. It adds 11 lines to the 80-line foreign budget and adds or removes no step, so no census re-pin is needed.

diff --git a/.github/workflows/t27b-native.yml b/.github/workflows/t27b-native.yml
index 75bad76bf..ad2de7558 100644
--- a/.github/workflows/t27b-native.yml
+++ b/.github/workflows/t27b-native.yml
@@ -30,15 +30,14 @@
 # from the lab's last cache artifact. The lab does serve /refcache.json, but it
 # is not that artifact: it is lab.py's own JSON, keyed by sha256 over the lab's
 # x86_64 t27c, while `t27b corpus --reference-cache` reads tab-separated rows
-# keyed by FNV-1a over the t27c binary that runs here (blockers.rs
-# toolchain_stamp). No lab row can ever hit. So the reference is built here
-# (t27c release + zig 0.16.0, natively) and its cache is carried between runs
-# with actions/cache: master's runs seed it, a pull request restores master's,
-# and only specs whose `use` closure changed are re-run. That holds only while
-# the t27c bytes hold: t27c embeds the commit, so the binary itself is cached,
-# keyed by the tree of bootstrap/ + Cargo.lock + Cargo.toml (#7548). A PR that
-# changes those rebuilds t27c, every key misses, and that run pays the full
-# reference sweep.
+# keyed by specs/tri/t27b/verdict_key.t27 (#8176): the spec and its use
+# closure, the code `t27c gen` prints for it, the test runner, zig and the
+# host. The lab is x86_64-linux and this job aarch64-linux, so no lab row can
+# hit. So the reference is built here (t27c release + zig 0.16.0, natively)
+# and its cache is carried between runs with actions/cache: master's runs seed
+# it, a pull request restores master's, and only specs whose key moved are
+# re-run. A PR that rebuilds t27c re-runs only the specs whose generated code
+# changed.
 #
 # No secrets: the repository is public; nothing here writes to GitHub.
 
@@ -282,7 +281,8 @@ jobs:
           files=$(jq '.results | length' run.json)
           mism=$(jq '.totals.jit_interp_mismatch' run.json)
           dis=$(jq '.totals.reference_disagree' run.json)
-          line="t27b $n / reference $d native arm64 (+$v pass_vacuous, $o outside the reference; $files files; jit/interp mismatch $mism; reference disagree $dis); ratchet ${{ steps.ratchet.outputs.verdict }}; corpus ${{ steps.corpus.outputs.seconds }} s"
+          reused=$(jq '.totals.reference.reused' run.json)
+          line="t27b $n / reference $d native arm64 (+$v pass_vacuous, $o outside the reference; $files files; jit/interp mismatch $mism; reference disagree $dis; reference reused $reused of $files); ratchet ${{ steps.ratchet.outputs.verdict }}; corpus ${{ steps.corpus.outputs.seconds }} s"
           echo "::notice title=t27b native::$line"
           {
             echo "### t27b native corpus (ubuntu-24.04-arm)"

t27b: key reference verdicts by the generated code, not the t27c bina…

Navigation Menu

Repository navigation

t27b: key reference verdicts by the generated code, not the t27c binary #8176

Boundary

User Scenarios

Requirements

Success Criteria

Requirements

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

t27b: key reference verdicts by the generated code, not the t27c binary #8176

Description

Boundary

User Scenarios

Requirements

Success Criteria

Requirements

Activity

gHashTag commented on Oct 9, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions