The census counted parse failures it never had

[gHashTag]

tri unparsed called every non-zero t27c check "the compiler cannot read".
Of 97 such specs, only 79 fail at parse:

stage	count
parse	79	the census operates here
typecheck	13	they parse perfectly
lex	4	unterminated string literal
semantic	1	nested fn captures enclosing locals

That is a category error, not a rounding difference: the census read a
construct off the failing line of a type error. specs/numeric/gf8.t27
stops at exp = exp + 1; — an assignment that compiles — rejected as
cannot assign F64 to F32. Seven gf* specs sat in the residue on exactly that
line, and the residue was the census's own confusion.

The discriminator is checked both ways: no typecheck output contains a parse
word, and no parse output contains Typecheck.

Five constructs added, each with probe and counter

row	probe (rejected)	counter (compiles)
&[_]T{} anonymous array literal	&[_][]const u8{}	[ 1, 2 ]
;; statement terminated twice	print("hi");;	print("hi");
}; a brace with nothing open	trailing };	no trailing brace
pub use NAME;	pub use session_timeout;	use session_timeout;
fn(a: T) R { } anonymous fn literal	const h = fn(e: E) void { };	const h = handler;

  blind spot          32 -> 10
  construct NAMED     43 -> 53
  counters compiling  16 of 16

is_assignment joins the abstention, justified by its own probe: x = x + 1;
compiles, so a failing line that is one is a symptom.

Two rows were NOT shipped, and both refusals are the point

Prose. A paragraph where a declaration belongs does stop the parser — its
probe is rejected — so a row looked obvious. Then the control:

loosest rule     fired on 8925 lines inside specs that PARSE
+ no brackets    97
+ sentence-like  42, and it lost 2 of the 5 real cases

Tuning was not the problem. Acceptance is position-dependent:
specs/api/sdk_contract.t27 parses while carrying

fn random(dim: usize, seed: u64) -> Hypervector
    Create random hypervector

so the same words after a body-less signature are fine. No line-level matcher
is faithful to that at any threshold. The row came out, the reason went into the
source, and the six specs went back into the blind spot — where
tri prose report, which asks the compiler line by line, already answers them.

Shipping it would have made the residue read 5 instead of 10.

A switch prong. .module => "module", looks like an easy row. Every probe I
could write fails — including the form copied verbatim from a spec that
parses, because in isolation it needs context the snippet does not carry. No
honest probe, therefore no row.

The census now

  specs tracked                       745
  ... refused at PARSE                 79   <- this census
  ... refused at TYPECHECK             13
  ... refused at LEX                    4
  ... refused on SEMANTICS              1
  ... construct NAMED and PROBED       53
  ... cause is UPSTREAM, not named     16
  ... not decided, nothing claimed     10
  broken ON PURPOSE under fixtures/    21
  constructs the compiler now ACCEPTS   2

53 + 16 + 10 = 79. The parts sum to the stage they belong to.
--list now prints the undecided lines, so the blind spot is visible rather
than carried.

Gates, run locally

specs generate 0 · specs parse 0 · conflict markers 0 · seals fresh 0 ·
types ratchet 0 · skill check 0 · prose report 0 · unparsed probe 0 ·
cargo test -p tri 0 (17 tests in unparsed.rs) ·
rustfmt --check cli/tri/src/unparsed.rs 0.

bootstrap/src/compiler.rs is untouched, so FROZEN_HASH does not move.

Refs #2864

[github-actions]

📓 NotebookLM Notebook linked to this PR

• Notebook ID: b83263109fb055dc
• URL: https://notebooklm.google.com/notebook/b83263109fb055dc

This notebook contains session context, decisions, and artifacts for this work.

[github-actions]

PR Dashboard

Generated at: 2026-08-29 21:17:06 UTC

• The census counted parse failures it never had #2886 The census counted parse failures it never had | Created: 2026-08-29 | Updated: 2026-08-29 | 20 failing / 25 total
• Land the formal verification harness on master (it has never run) #2727 Land the formal verification harness on master (it has never run) | Created: 2026-08-27 | Updated: 2026-08-27 | 5 failing / 24 total
• chore(deps)(deps): bump thiserror from 2.0.18 to 2.0.20 #2693 chore(deps)(deps): bump thiserror from 2.0.18 to 2.0.20 | Created: 2026-08-24 | Updated: 2026-08-28 | 3 failing / 34 total
• chore(deps)(deps): bump hyper from 1.10.1 to 1.11.0 #2692 chore(deps)(deps): bump hyper from 1.10.1 to 1.11.0 | Created: 2026-08-24 | Updated: 2026-08-28 | 3 failing / 33 total
• chore(deps)(deps): bump base64 from 0.22.1 to 0.23.1 #2691 chore(deps)(deps): bump base64 from 0.22.1 to 0.23.1 | Created: 2026-08-24 | Updated: 2026-08-28 | 3 failing / 34 total
• chore(deps)(deps): bump serde from 1.0.228 to 1.0.229 #2690 chore(deps)(deps): bump serde from 1.0.228 to 1.0.229 | Created: 2026-08-24 | Updated: 2026-08-28 | 3 failing / 34 total
• chore(deps)(deps): bump uuid from 1.23.2 to 1.25.0 #2689 chore(deps)(deps): bump uuid from 1.23.2 to 1.25.0 | Created: 2026-08-24 | Updated: 2026-08-28 | 3 failing / 34 total
• Unbreak FPGA CI and correct two unbacked hardware claims #2081 Unbreak FPGA CI and correct two unbacked hardware claims | Created: 2026-08-11 | Updated: 2026-08-20 | 5 failing / 27 total

Summary

Status	Count
Total Open PRs	8
PRs with Failing Checks	7
PRs with All Checks Green	1
READY	0
FAILING	7
PENDING	0

Seal Status

• ⚠️ STALE -- sha256(compiler.rs)=f569da0c6c14 != manifest seal=87e5cbd3ad94.
  The committed NMSE numbers were certified against an older compiler.rs.
  Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).