[EPIC] t27b for t27o: every spec t27o emits runs natively under t27b and agrees with t27c (100 %) #10879

Activity

1. gHashTag

   added

   phi-loopPHI LOOP development cyclePHI LOOP development cycle

   tooling

   performancePerformance-related issuePerformance-related issue

   epic

   on Oct 19, 2026
2. gHashTag commented on Oct 19, 2026

   

   gHashTag

   on Oct 19, 2026

   Last edited by gHashTag

   OwnerAuthor

   More actions

   Step 1 measured (master 57aa3bc; data: the lane's RESULTS.txt)

   Corpus: every t27o output with at least one converted fn or test -- Rust: 492 specs from all 1,388 tracked .rs files of the five roadmap repositories; Zig: 724 specs from trinity src/ + trios crates/ (#10819).

   	specs	t27b pass (asserts / 0 asserts)	fail	unsupported	front end	codegen / JIT mismatch
   Rust	492	387 (146 / 241)	2	97	6	0 / 0
   Zig	724	706 (560 / 146)	9	9	0	0 / 0
   Total	1,216	1,093 (89.9 %)	11	106	6	0 / 0

   • Agreement: seeded sample of 150 (seed 10879): 0 disagreements between t27b and t27c in 135 compared files.
   • The 123 not passing, under the reference: 58 pass in t27c -- the real t27b gap (53 Rust, 5 Zig); 55 do not compile in t27c either -- converter defects, filed as [bug] t27o emits 55 specs that t27c cannot compile either -- undeclared names, negative enum(u32) values, _ identifiers, >>=, open ranges #10905; 10 fail in both, on the same tests.
   • The gap is mostly eagerness: in 53 of the 58, every blocker sits in a fn or module const that no test reaches. t27c's Zig backend never analyzes those; t27b lowers them eagerly. The mechanism exists for signatures (lazy_sig_plan.t27), not for bodies or consts.
   • Blocker issues, in order: t27b for t27o: ExprIdentifier(undeclared) -- unlocks 24 of 1216 t27o specs (#10879) #10897 (unreached bodies: ExprIdentifier undeclared, +37 specs; first PR, the others lean on it), t27b for t27o: ConstDecl(slice) -- unlocks 9 of 1216 t27o specs (#10879) #10899 (unreferenced slice consts), t27b for t27o: ConstDecl -- unlocks 2 of 1216 t27o specs (#10879) #10903 (unreferenced consts), t27b for t27o: type mismatch -- unlocks 8 of 1216 t27o specs (#10879) #10901 (type mismatch in unanalyzed code), t27b for t27o: ExprArrayLiteral -- unlocks 8 of 1216 t27o specs (#10879) #10900 (array literal, a REPEAT form).
   • 0-assert passes: 387 of the 1,093 passes run no assert -- the converted spec has fns but no tests. For t27o that is the oracle's job (--check adds oracle tests); the acceptance metric counts them, the golden metric must not.
   • Side finding: use tri::crypto::sha256 (or hex, sha512) next to use tri::t27b::type_text breaks use resolution in t27c too -- both declare same (t27c gen: two imported modules that each declare a helper of the same name make the splicer drop both, so a pulled function loses its own helper #10556); tri t27o --check cannot pass bootstrap/src/compiler.rs or main.rs today under either runner.
3. gHashTag commented on Oct 19, 2026

   

   gHashTag

   on Oct 19, 2026

   OwnerAuthor

   More actions

   Progress 2026-10-19 evening.

   Merged:

   • t27b: do not lower bodies no test reaches -- Sema-level refusals in an unreached fn are withdrawn from one plan (Closes #10897) #10999 closes t27b for t27o: ExprIdentifier(undeclared) -- unlocks 24 of 1216 t27o specs (#10879) #10897: unreached fn bodies. t27b runs 1,130 of the 1,216 t27o specs (92.9%, from 89.9%); the real gap (t27c passes, t27b refuses) went from 58 to 21 specs. All 37 newly passing t27o specs agree with t27c. cargo test -p t27b green on linux and macos at the merged head d11f776; hand lines in cli/t27b net -9. Ledger rows: parser and deps_card pass, notebooklm's blocker is BenchBlock, cap 129 -> 127.
   • t27c: the use splice reads a name inside a pulled declaration in that declaration's module (Closes #10556) #11012 closes t27c gen: two imported modules that each declare a helper of the same name make the splicer drop both, so a pulled function loses its own helper #10556: the use splice reads a name inside a pulled declaration in that declaration's module (specs/compiler/use_splice.t27, net -112 hand lines). This is the same / type_text blocker named in the side finding; tri t27o --check on compiler.rs goes from BLOCKED to t27c 87/87.

   Open, next lane (TB2, running): #10899 and #10903 (unreferenced module constants, the constant variant of the #10897 rule), then #10901 (type mismatch in an unreached fn) and #10900 (array literals).

   Note: these four issues had been closed as "not planned" by the issue-shape gate at 06:38Z and were reopened by the self-heal of #11007; the gate itself is fixed by #11008.
4. gHashTag commented on Oct 19, 2026

   

   gHashTag

   on Oct 19, 2026

   Last edited by gHashTag

   OwnerAuthor

   More actions

   Progress 2026-10-19, late.

   Merged: #11046 (closes #10899, #10903 and #11053). A module constant nothing analyzed reaches is lazy like an unreached body; the reach (analyzed in ast_shape.t27) now covers constants and subsumes two hand-written exemptions (cli/t27b hand lines +27 -28, net -1). The same PR closes the hole in #10999: a local of a type declared nowhere, in an unreached fn, keeps its refusal.

   Measured (t27b built from the branch against the master build, --jobs 1): Rust 421 -> 434, Zig 710 -> 710, total 1,131 -> 1,144 of 1,216 (94.1%), unsupported 68 -> 55; 13 specs gained, none lost, no per-test verdict changed; --reference t27c over the 13 and over the two newly decided tracked specs: 0 disagreements. cargo test -p t27b green on linux and macos; ledger rows recounted in the PR (specs/lsp/schema.t27 pass, cap 127).

   Left on this epic: #10901 (type mismatch in an unreached fn, +8 specs in the first measurement) and #10900 (array literals, +8); the 55 specs t27c cannot compile were made valid or refused by #10976; the real gap (t27c passes, t27b refuses) was 21 specs after #10999 and has not been re-measured since #11046; the next lane runs t27b corpus --blockers first.