Corpus census W737-W739: the defect population is 168, not 490 — and what is still open

[gHashTag]

What this issue preserves

Waves W729–W739 worked the compiler and the corpus census. The compiler changes are landed; this issue holds the measurements and the open work, so the next wave starts from the numbers instead of re-deriving them.

The honest corpus, measured on one pinned binary

t27c classify first (a .t27 extension is a filename, not a type declaration):

class	files	share
SOURCE module …	590	95.5%
NOT-CODE (Markdown document)	14	2.3%
ALT-SYNTAX spec X { … }	8	1.3%
UNCLASSIFIED	5	0.8%
MIXED	1	0.2%
total	618

Then test-report over all 578 sources, binary hash 56b7a6b6d8047566 recorded before and after:

run their tests        88     1293 tests passing
BLOCKED               490

And every blocked spec crossed with spec-status:

status	specs	meaning
IMPLEMENTED	168	written and genuinely failing — the defect population
UNWRITTEN	154	no bodies; blocking is correct
NOPARSE	144	does not parse
NOFN	18	no functions at all
PARTIAL	6

The real defect population is 168, not 490.

Three ways the denominator was wrong

1. Files that are not source — 28 of 618 (t27c classify).
2. Sampled by alphabetical prefix — 131 specs gave 26.7% running; all 578 give 15.2%. specs/a*…specs/f* holds the hand-maintained numeric and fpga modules.
3. Unwritten counted as broken — all 101 specs of unable to format type '@TypeOf(undefined)' contain @compileError("not yet implemented"). Fixed in W739; they now report honestly.

Compiler changes landed

wave	change	effect
W729	type alias const X = u8 no longer emitted as @"u8"	unblocked the alphabet SSOT
W730	primitives never escaped as identifiers	@"prim" 8 → 0 of 130
W730	bench-path double discard suppressed
W732	t27's bare slice [T] → Zig []T	error class removed
W734	for x in collection (was range-only)	4 of 5 specs parse
W735	parameter shadowing a module declaration	+59 tests (router 19, testbench 16, timing 24)
W736	local shadowing, fn bodies only	+15 tests (memory)
W739	stop formatting undefined	101 specs report their true status

Reference suite held at 145 tests through every change.

Open, with roots named

• undeclared identifier 'assert' — 60 specs. Third census class, never worked to root. May well be the same disease as the 101 (status, not defect).
• Local shadowing in test/bench blocks — 9 specs. A test block's first assignment parses as StmtAssign, not StmtLocal; extending the rename renamed references while the binding kept its old name, turning a shadow error into use of undeclared identifier. Attempted and reverted — see collect_shadowing_locals' doc comment.
• if as an EXPRESSION without parentheses. parse_if_stmt has accepted it since W578; parse_if_expr still demands parens. Mirroring it — including the no_struct_literal guard — cost 46 of 145 reference tests, because a branch value may legitimately be a struct literal where a statement body may not. Attempted and reverted.
• Grammar gaps still open: newtype struct X(T);, while (c) : (step), -> &str, module a::b, open slices arr[i..].
• The 168 written-and-failing specs have never been decomposed by cause.

Protocol notes worth keeping

• Never rebuild the compiler under a running census. The first full run had it rebuilt twice beneath it; 272 rows were measured across three binaries and the run was discarded, not salvaged. Pin the binary hash before and after.
• Classify by intent before symptom. t27c spec-status answers UNWRITTEN/PARTIAL/IMPLEMENTED/NOPARSE/NOFN and did so all along.

Data: docs/reports/CORPUS-CENSUS-W737.tsv, docs/reports/CORPUS-BLOCKED-BY-STATUS-W738.tsv. Theorems T255–T276 in docs/theory/IGLA-FORMAL-RESULTS.md.

[gHashTag]

A boundary drafted from the paths this issue already names, so a bee can take it. Nothing is reserved until this is in the issue body - the Queen reads the body, not comments. Check the paths first: a wrong boundary reserves files the work does not own.

Boundary

• docs/reports/CORPUS-BLOCKED-BY-STATUS-W738.tsv
• docs/reports/CORPUS-CENSUS-W737.tsv
• docs/theory/IGLA-FORMAL-RESULTS.md