[shd101wyy]

The seed wait is satisfied. v0.2.65 is published and carries #1438: a stage 1 built by the v0.2.65 seed no longer crashes on move_only (the v0.2.64-seed crash — a lost E0901 inside a Box(sub) arm of _augment_arm_pattern — is the #1438 bug and is fixed by the seed). Re-gated and marked ready on the v0.2.65 seed; see the re-gate block below.

Summary

VBD V4 (plans/VALUES_BY_DEFAULT.md §6 V4, decisions 21 and 31): Pattern (src/pattern.yo) becomes a value enum with unique Box children. It follows #1435 (Z3Sexpr) and #1438, the unwind S1 that this conversion exposed.

• Representation. At's and ThroughBox's sub are Box(Self); the sub-pattern lists stay ArrayList(Self).
• Clone. derive(Clone) refuses an enum that mixes Box(Self) with ArrayList(Self), so Pattern has a manual Clone. It deep-clones through Box.clone and clone_patterns, and shares the field_idx / elem_tys handles.
• Arms and constructions. Arms spell Box(sub), which the seed carries. Constructions use box(...).
• Element access (the ArrayList is still a handle). For a move-only element, get, .clone() of the list, for(&xs, …) and match(&xs(i), …) are E0901. The tree therefore reads elements as &subs(i) arguments and subs(i).clone().
  • The usefulness walk's row specialization moved into per-row helpers: _specialize_variant_row, _specialize_aggregate_row, _expand_row_head and the _head_* predicates.
  • _useful_at delegates to _useful_head.
  • These touch src/pattern.yo, src/evaluator/exprs/match.yo, src/evaluator/exprs/pattern_compile.yo, src/codegen/exprs/match.yo and src/verifier/vc.yo.
• New issue (S2, not fixed here): issues/2026-10-17-a-match-through-a-borrowed-element-copies-the-element.md. match(&xs(i), …) is E0901 even though E0901 names that form as its repair.
• Docs. .github/instructions/yo-syntax.instructions.md gets the Box-enum pitfalls, and the plan gets V4's "as built" note.

Verification

Tier 1 re-gate on the v0.2.65 seed (2026-10-18; first at develop 5eed91253 / head e4473ee22, then rebased clean over develop 057420e4a — #1445, #1447's match-trap hardening, #1434, #1446, #1430 — to head 661efba12, full re-gate re-run):

• Stage 1 yo build --std-path ./std with the v0.2.65 seed: rc=0 — the step that crashed on the v0.2.64 seed. The first re-gate needed --cc zig (the box's GCC 15.2 rejected the then-seed's exhaustive-enum-switch emission under yo's -Werror=return-type, identically on clean develop); after fix(codegen): a statement-position match in a non-void function no longer falls off the end under gcc (S1) #1447's trap closers merged, the plain default-cc (gcc) build is rc=0 — the workaround is retired.

• The fix(codegen): a statement-position match in a non-void function no longer falls off the end under gcc (S1) #1447 interaction: this branch's codegen changes (the pattern walks in _collect_place_names / _collect_bind_decls / _emit_pattern_chain / _lookthrough_switch_cell, lines ~121-2300) and fix(codegen): a statement-position match in a non-void function no longer falls off the end under gcc (S1) #1447's construct-level trap closer in generate_match_expression (~2480) are disjoint; the trap keys on the evaluator's control_flow stamping, which this branch does not touch (0 references). match_nested — the file fix(codegen): a statement-position match in a non-void function no longer falls off the end under gcc (S1) #1447 extended with trap tests — passes 28/28 through this branch's Pattern lowering.

• Stage 2 (the binary stage 1 produced): check ./src rc=0; check ./std rc=0.

• Feature tests, one file per call, --parallel 1 (head 661efba12): move_only 55/55, match_nested 28/28 (25 pre-fix(codegen): a statement-position match in a non-void function no longer falls off the end under gcc (S1) #1447 + its 3 new), match_at_binding 4/4, algebraic_effects 79/79; internal/match_binding_shadow 3/3 in CI mode (YO_TEST_LEAK_VERDICT=0), the file fix(codegen): a statement-position match in a non-void function no longer falls off the end under gcc (S1) #1447 names.

• yo fmt --check with the tree-built binary on every changed file: rc=0.

• scripts/check-issue-refs.sh: rc=0, after repointing one citation the branch had added to its moved home (issues/fixed/a-hashmap-of-a-move-only-key-or-value-does-not-compile.md).

• Leak comparison against develop's binary (YO_TEST_LEAK_VERDICT=1, ASan+LSan test binaries, branch stage 2 vs develop 5eed91253 built the same way, same test files; measured at head e4473ee22 — the develop move to 057420e4a touches none of the three files, so the table stands):

  tests/internal file	this branch (e4473ee22)	develop (5eed91253)
  verifier_quantifiers	4 passed, 2 failed — both LeakSanitizer-only, each 22,549 B / 336 allocs	identical: same 2 tests, same signature
  module_invalidation	1 passed, 11 failed — all LeakSanitizer-only: 9 × 22,549 B / 336, 1 × 22,933 B / 342, 1 × 23,125 B / 345	identical: same 11 tests, same three signatures
  verifier_loops	12 passed, 2 failed — both LeakSanitizer-only, each 22,549 B / 336 allocs	identical: same 2 tests, same signature

  Verdict: identical leak counts and byte/allocation signatures on every file — the "identical or fewer" criterion holds; every leak pre-exists on develop. (Method: test binaries built with gcc + -fsanitize=address through a wrapper appending -Wno-error=return-type, which only downgrades the GCC-15-only switch-fallthrough severity back to the warning CI's compilers emit; the LSan verdict was verified live — it fails leaking tests and prints full reports. module_invalidation is 12/12 in CI mode (YO_TEST_LEAK_VERDICT=0), so all 11 failures are verdict-only.)

Earlier A/B, same-input against develop's binary (both built from 5425cd7bf with seed v0.2.63, both stage-2 binaries linking mimalloc):

• Emit-diff (compile src/main.yo --emit-c --skip-c-compiler): 0 lines.
• Allocation ratchet, check src/main.yo: 1,828,286,478 calls against develop's 1,823,309,567 (+0.27%, from the row and arm-pattern clones). That is within ±5% of the baseline 1,799,052,684.
• yo verify ./tests/spec --format json: identical output.

Tier 1 on the earlier rebased tree (develop fbd79fc90 plus this branch, stage 1 built with seed v0.2.64 — superseded by the re-gate above; kept as the record):

• yo build --std-path ./std: rc=0
• check src / check std: rc=0 / rc=0
• Tests (one file per call): the stage 1 is built by the v0.2.64 seed:
  • passed: match_nested 25, match_or_variants 6, match_at_binding 4, match_structs 9, gadts 10, recursive_enum 6, spec/verify_match_patterns 6, internal/verifier 29, internal/match_binding_shadow 3, rc 114;
  • internal/verifier_match: 1 failure, a LeakSanitizer report only (see the leak comparison below);
  • move_only: the batch crashes, because this stage 1 is compiled by a seed without fix(codegen): an unwind through a frame whose destructor makes a call is no longer lost (S1) #1438 (see the banner);
  • stage 2 (built by that stage 1, which carries fix(codegen): an unwind through a frame whose destructor makes a call is no longer lost (S1) #1438): the stage-2 build is rc=0. The repro (match(o, (w := .Some(f)) => …) on a move-only Option) now reports its E0901 instead of SIGSEGV. move_only passed 55, match_nested 25, match_at_binding 4 and algebraic_effects 79.
• Earlier run on 5425cd7bf (pat-S1, before fix(codegen): an unwind through a frame whose destructor makes a call is no longer lost (S1) #1438), 29 files:
  • 24 passed;
  • move_only crashed: that was the unwind S1, fixed by fix(codegen): an unwind through a frame whose destructor makes a call is no longer lost (S1) #1438, and it is in the gate above;
  • 4 tests/internal files (verifier_quantifiers, module_invalidation, verifier_loops, verifier_match) failed on LeakSanitizer reports only. CI runs these with YO_TEST_LEAK_VERDICT=0.
• Leak comparison, tests/internal/verifier_match.test.yo, solver-free datatype test: develop's binary (the 63a43b8f8 tree and its stage 1) fails the same one test with the same report, 22,805 bytes in 340 allocations, identical to this branch's. It predates this change.
• yo fmt --check with the tree-built binary on every changed file
• scripts/check-issue-refs.sh on the final rebased head
• Plan progress header updated
• Docs in docs/: not applicable

Timing. The check ./src time and stage-2 RSS A/B (exclusive, medians of 3) is pending, measured after merge. It was cancelled at the coordinator's request while the heavy queue was 35 deep. Given a byte-identical emission and a +0.27% allocation delta, no measurable change is expected.

LLM attribution

• Model: Claude Opus 5.5
• Agent harness (optional): Claude Code
• v0.2.65 re-gate, leak comparison and ready-staging (2026-10-18): GLM-5.3, ZCode harness

Deferred

• VcSort / VcTerm wait on issues/questions/2026-10-17-should-the-verifier-terms-take-box-or-rc-children.md.
• The match(&xs(i), …) S2 above is not fixed here.
• The timing A/B is measured after merge.

🤖 Generated with Claude Code