Repository navigation
[GOLD-RING] parser: the Zig subset lexer.t27 needed (Closes #2426) - #2427
Conversation
The seal move here is made on the Architect's explicit authorisation, given in answer to #2368's last section, which asked for exactly that before touching bootstrap/src/compiler.rs. 12 specs move from the compile-debt ledger to generating, with zero regressions: before: 712 specs, 541 generate, 171 known-broken after: 712 specs, 553 generate, 159 known-broken The ledger is updated in the same commit and by the same compiler that changed, so the gate now HOLDS those twelve compiling rather than excusing them. Measuring a ledger with the compiler it gates is the lesson of #2365; this is the first change written under it. Five gaps, each verified by the error moving rather than by hope: 1. the anonymous braced literal .{} / .{ a, b } / .{ .f = a } 2. an anonymous struct TYPE in return position 3. Zig's slice spellings x[a..b] and the open-ended x[a..] 4. keyword-named enum values, `return .const;` 5. the struct-method skip mistaking a return type's brace for the body's Gap 5 is why the campaign was worth running. Struct methods are skipped brace-balanced and the skip stopped at the first brace; with an anonymous struct return type that brace belongs to the TYPE, so the balance closed on the type, the struct ended at its first such method, and the module ended at the struct's own brace. compiler/parser/lexer.t27 lost everything after line 362 while reporting a parse error 144 lines further on. The code already carried a comment about this exact hazard -- W577, which cost specs/jit/jit.t27 797 of its 875 lines the same way. The hazard was documented and recurred on a construct the comment did not anticipate. Two corrections to my own earlier reports. I wrote in #2368 that line 362 parses while line 424 does not; NEITHER parsed -- the first was consumed by the skip and looked like success, which is the failure mode being fixed. And gap 3's fix went into the slice parser first and changed nothing, because `..` sits in the binary-operator chain and was eaten before the slice code saw it. Ceremony per FROZEN.md section 5: frozen-digest re-run, seal line replaced, cargo build --release green. Resealing invalidates the NMSE certification in repro/numerics/nmse_manifest.json -- nmse_gf16.py --seal needs re-running after refreezing, and that is left to the owner. Six specs remain, each for its own reason, exactly as #2372 triaged them. This campaign closed the family lexer.t27 needed and no more.
# Conflicts: # bootstrap/stage0/FROZEN_HASH
|
Rebased onto current master (seal conflict only — recomputed from the merged |
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
PR DashboardGenerated at: 2026-08-22 15:08:49 UTC
Summary
Seal Status
|
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
PR DashboardGenerated at: 2026-08-22 15:27:23 UTC
Summary
Seal Status
|
…2426) Wave 2 of the campaign authorised in #2426. 541 generate -> 558, ledger 171 -> 154, zero regressions. Four gaps, each verified by the error moving: 6. keyword-named enum MEMBERS -- then, and, assert, expect are real members of ClauseKind and requiring Ident dropped them 7. a nested type declaration inside a struct body, const X = enum {...}; 8. payload capture in if -- if (opt) |value| { ... } 9. Zig's while continue expression -- while (i < n) : (i += 1) { ... } Gap 7 is the THIRD appearance of one shape: a nested brace group ending the construct that contains it. W577 found it for methods, wave 1 found it for an anonymous struct return type, and this is the same failure for a nested type -- the struct body loop stops at any RBrace, so the enum's closing brace ended the struct and the module ended at the struct's own. Gap 9 attaches the continue expression rather than discarding it. Dropping it would leave a loop whose counter never advances, which compiles and never terminates; refusing the file is better than emitting that silence. compiler/codegen/verilog/codegen.t27 now generates, joining lexer.t27.
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
|
Wave 2 — four more gaps, and the totals grow. Seventeen specs leave the compile-debt ledger, zero regressions, and
Gap 7 is the third appearance of one shape. A nested brace group ending the construct that contains it: W577 found it for methods, wave 1 found it for an anonymous struct return type, and this is the same failure for a nested type — the struct-body loop stops at any Gap 9 attaches the continue expression rather than discarding it. Dropping it would leave a loop whose counter never advances — code that compiles and never terminates. Refusing the file is better than emitting that silence, which is the same reasoning as the open-ended slice getting its own marker in wave 1. Still six specs short of the file being clean, and each remaining one has its own distinct cause: |
#2426) Wave 3 of the campaign authorised in #2426. 541 generate -> 562, ledger 171 -> 150 across three waves, zero regressions. 10. tagged unions: union(enum) { A: struct {...}, ... } 11. a call on an ARBITRARY expression: func.?(a, b) 12. the inline unroll hint before for / while union is not a keyword in this lexer, so union(enum) parsed as a CALL whose argument is the keyword enum. Captured verbatim rather than lowered to a struct: its variants carry payload types, and emitting a union as a struct would give every variant its own storage and silently change the layout. The expression-call branch replaces a comment reading 'shouldn't normally happen here'. specs/jit/jit.t27 is exactly it -- the optional unwrap parsed, and the call on its result was left unconsumed, so the statement parser met a bare '(' and pointed past the construct rather than at it. inline is likewise not a keyword, so 'inline for' parsed as the expression 'inline' followed by a loop keyword. Skipped, not recorded: a parser that does not unroll loses nothing by ignoring the hint. compiler/codegen/zig/codegen.t27, specs/jit/jit.t27 and specs/ternary/bigint.t27 now generate. Five of the seven target specs are done; main_zig_handwritten needs an error-union return type, and cap_test is the compiler deliberately refusing closures, not a gap in the grammar.
|
Wave 3 — the campaign's totals, and five of seven target specs done. Twenty-one specs leave the compile-debt ledger, zero regressions at every step.
Gap 11 replaces a comment that read "shouldn't normally happen here." Where the seven stand
The last one is worth stating plainly: it is a wanted feature written as a passing test, and it belongs in the ledger until closures exist. No amount of parser work will move it. |
FROZEN_HASH was the only conflict, and it is the one kind of conflict that must not be resolved by choosing a side. Each candidate hashes its own branch's compiler.rs; after the merge neither describes the file that exists: ours 8e62cacb81c6e84d theirs 4f003654a44a4348 merged 6e2bad56817414a6 Recomputed from the merged bytes. `bootstrap/build.rs` verifies it on every build, so a picked side would have failed loudly -- but the reflex to pick one is wrong and worth naming. The corpus ratchet then failed for the good reason: this branch FIXES parse for 17 specs the ledger still lists as broken. Five of those seventeen reach the next check and fail `parse-no-discard` instead -- the same specs, one stage deeper, not new damage. Ledger 220 -> 208: seventeen removed, five re-entered at the phase they now reach, with the reason saying which. check_specs_generate goes 541 -> 562. Elaboration 176, baseline 176, clean. Verilog widths 2 known, 0 new. Refs #2427
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
PR DashboardGenerated at: 2026-08-23 16:45:28 UTC
Summary
Seal Status
|
`bootstrap/stage0/FROZEN_HASH` holds sha256 of the file it seals. When two branches both change `compiler.rs`, git conflicts the seal and offers two hashes, and BOTH are wrong -- each describes its own side's file, and the merged file is a third thing: ours 8e62cacb81c6e84d theirs 4f003654a44a4348 merged 6e2bad56817414a6 Every reflex a conflict trains -- take ours, take theirs, pick the newer -- writes a seal for a file that does not exist. `build.rs` catches it, so the cost is confusion rather than corruption, but the command exists to remove the temptation. `tri reseal check` reports three states, not two: matching, mismatched, and CONFLICTED. The third needs its own message because a conflicted file trimmed into a comparison reports a plain mismatch against `<<<<<<< HEAD` and sends the reader to fix the wrong thing. `tri reseal write` discards both candidates and recomputes. Written through a temporary and renamed: a seal truncated by an interrupted write agrees with nothing, and reads exactly like a seal that was never updated. Verified in all four states on a real repository -- matching, deliberately mismatched, conflicted, and conflicted-then-written -- with the seal returned to its original value afterwards. Three unit tests cover the conflict recognition, including what `trim()` would have produced. Skill sections 78 and 79: a derived value in conflict has no correct side, and a ratchet that fails because you fixed something. Refs #2427
The gate that would have caught #2427. $ tri mods orphan --gate # at origin/master ::error::cli/tri: orphaned rose 0 -> 1. A file left the build and nothing else would have said so. exit 1 $ tri mods orphan --gate # after the fix in this branch ORPHAN CEILING: CLEAN exit 0 Exact match in both directions, following MAX_SORRY: a count above its ceiling is a regression, and a count below it is slack the next orphan can hide in. Both were exercised -- a planted file in cli/tri/src exits 1, and a ceiling raised to 3 against an actual 0 exits 1 with the instruction to lower it. `--self-check` is the negative control. It builds a two-file crate in a temp directory, one declared and one not, and exits non-zero unless exactly the undeclared file comes back -- so a green run from this gate means the walk was working, not that it was skipped. The workflow carries NO `paths:` filter, and the comment says why: the change that strands a file is a change to the crate ROOT, not to the file that goes missing. A filter on the stranded path is silent for exactly this defect. bootstrap's 7 orphans are recorded, not repaired. `tooling.rs` alone holds 7 gate bodies (#2905), and connecting any of them is a decision about what goes red rather than a repair. The ledger says so in the file, beside the number. Refs #2900 Refs #2905
…ilent (#2911) * fix(tri): restore `mod elab;` -- the command has not existed since #2427 d646071 ("[GOLD-RING] parser: the Zig subset lexer.t27 needed", a change to bootstrap/src/compiler.rs) removed three lines from cli/tri/src/main.rs in one hunk: the `mod` declaration, the `Elab` variant and the dispatch arm. The 319-line file stayed. Since then: $ tri elab classify error: unrecognized subcommand 'elab' Nothing said so. `cargo build` cannot fail on a file it does not compile, and the test suite went 358 -> 354 without a gate that reads the number. Four tests left the suite in silence. `tools/check_elab_ratchet.py` still instructs its reader to run this command, in its docstring and in its failure output, and two NOW documents describe it as working. `tri elab` was written because that same script once counted iverilog's per-file summary line `N error(s) during elaboration.` as a diagnostic -- 25 phantoms in a reported 186, quoted onward before anyone read the rows. Restored, it reads: real diagnostics: 176 summary lines NOT counted: 28 Nothing else changes: elab.rs is untouched, and its four tests pass as written. Closes #2900 * feat(tri): `mods orphan` -- source files no crate root reaches The general form of the defect in the parent commit. Rust compiles a file only if a `mod` declaration leads to it from the crate root; a file nothing declares is not unused, it is a file the compiler never opens. It cannot warn, cannot fail to build, and its `#[test]` functions do not exist for `cargo test`. bootstrap 97 files under src/, 90 compiled, 7 orphaned bootstrap/src/_enrich/mod.rs 441 lines bootstrap/src/math_pslq.rs 589 lines bootstrap/src/neural/attention.rs 4 lines bootstrap/src/neural/hslm.rs 4 lines bootstrap/src/runtime_minimal.rs 19 lines bootstrap/src/runtime_minimal_test.rs 18 lines bootstrap/src/tooling.rs 409 lines cli/tri 33 files under src/, 33 compiled, 0 orphaned Historical control: the same binary run against origin/master, where `mod elab;` is still missing, prints cli/tri 32 files under src/, 31 compiled, 1 orphaned cli/tri/src/elab.rs 319 lines 4 #[test] that cargo cannot see It fires on the defect and is silent after the fix. Resolution follows the language rather than a heuristic. `mod c;` in `a/b.rs` means `a/b/c.rs` or `a/b/c/mod.rs`, never `a/c.rs`; `#[path]` and `include!` are edges; inline `mod c { }` needs no file. The loose rule -- is this stem named by any `mod` anywhere -- hides a real orphan behind a same-named module in an unrelated directory, and its mutation fails the first test here. `reached` counts files outside `src/` too, because `#[path = "../../gen/..."]` is a real edge, so the report does not subtract the two sets and call the difference orphans; it prints the two comparable ones and says where the rest went. Refs #2900 * feat(ci): ratchet the orphaned-module count, with a negative control The gate that would have caught #2427. $ tri mods orphan --gate # at origin/master ::error::cli/tri: orphaned rose 0 -> 1. A file left the build and nothing else would have said so. exit 1 $ tri mods orphan --gate # after the fix two commits back ORPHAN CEILING: CLEAN exit 0 Exact match in both directions, following MAX_SORRY: a count above its ceiling is a regression, and a count below it is slack the next orphan can hide in. Both were exercised -- a planted file in cli/tri/src exits 1, and a ceiling raised to 3 against an actual 0 exits 1 with the instruction to lower it. `--self-check` is the negative control. It builds a two-file crate in a temp directory, one declared and one not, and exits non-zero unless exactly the undeclared file comes back -- so a green run means the walk was working, not that it was skipped. The workflow carries NO `paths:` filter, and the comment says why: the change that strands a file is a change to the crate ROOT, not to the file that goes missing. A filter on the stranded path is silent for exactly this defect. bootstrap's 7 orphans are recorded, not repaired. `tooling.rs` alone holds 7 gate bodies (#2905), and connecting any of them is a decision about what goes red rather than a repair. The ledger says so in the file, beside the number. Refs #2900 Refs #2905
[GOLD-RING] — the seal move in this pull request is made on the Architect's explicit authorisation, given in response to #2368's last section, which asked for exactly that before touching
bootstrap/src/compiler.rs.Result
12 specs move from the compile-debt ledger to generating, with zero regressions.
The ledger is updated in the same commit and with the same compiler that changed, so the gate now holds those twelve compiling rather than excusing them. Measuring a ledger with the compiler it gates is the lesson of #2365; this is the first change written under it.
Five gaps, each verified by the error moving
.{},.{ a, b },.{ .f = a }fn f(…) struct { u64, []const u8 }x[a..b]and the open-endedx[a..]return .const;,.use,.dataGap 5 is the reason the campaign was worth running. Struct methods are skipped brace-balanced, and the skip stopped at the first
{. With an anonymous struct return type that brace belongs to the type, so the balance closed on the type, the struct ended at its first such method, and the module ended at the struct's own}.compiler/parser/lexer.t27lost everything after line 362 while reporting a parse error 144 lines further on.The code already carried a comment about exactly this hazard — W577, which cost
specs/jit/jit.t27797 of its 875 lines the same way. The hazard was documented and recurred on a construct the comment did not anticipate.Two corrections to my own earlier reports
read_numberat line 362 "parses" while the two-parameter form at 424 does not. Neither parsed. The one-parameter form was consumed by the skip and looked like success — which is precisely the failure mode being fixed here. The bisect that found it tested prefixes of the struct, method by method, rather than trusting the reported line...is in the binary-operator chain and was eaten before the slice code ever saw it. The working fix leaves the..alone when the next token is]or).Ceremony
frozen-digestre-run and the seal line replaced, per FROZEN.md §5.cargo build --releasegreen. Resealing invalidates the NMSE certification inrepro/numerics/nmse_manifest.json;nmse_gf16.py --sealneeds re-running after refreezingstage0/FROZEN_HASH, and that is deliberately left to the owner rather than done here.What this does not fix
Six specs remain, each for its own reason, exactly as #2372 triaged them:
union(enum),func.?(),inline for, keywords as enum members, closures, and one corrupted-source case already repaired separately. This campaign closed the family thatlexer.t27needed and no more.Refs #2368, #2372.
🤖 Generated with Claude Code