EPIC: t27c silent misreads to zero and kept there -- lab on Railway, steward, follow-ups

[gHashTag]

Goal

Finish the t27c "parses, typechecks, generates, and is wrong" stack and keep it finished: every spec-shape pair tri misread --list reports is either refused by typecheck or fixed in the spec, and every generated line a backend emits compiles. This epic is the board card and the report channel for that work.

The owner delegated it on 2026-10-04: experiments run on Railway, not on the workstation (load average 600-840 that day; one cargo build --release -p t27c -p tri took 15 min 32 s locally, the same build takes 30 s on the lab), improvement and monitoring go through the Queen, and progress is reported here.

Map

What	Where
Lab (runs the compiler's CI gates on every watched head)	https://t27c-lab-production.up.railway.app/latest.json, one run: /runs/<sha>.json, one gate's log: /runs/<sha>/<gate>.log. Source infra/t27c-lab/lab.py. No secrets, GET only, never writes to GitHub
Steward (the profile agent that runs the loop)	.claude/agents/t27c-steward.md
Measure	tri misread --list (silent / refused spec-shape pairs), t27c suite --corpus-only --ratchet (ledger cap)

Stack (merge order; the owner merges)

• fix(t27c): refuse macro calls and typeless fields; rewrite eight Markdown specs (#5923) #5947 (Closes t27c misreads Rust macro calls and typeless fields; eight .t27 files are Markdown #5923) base master
• fix(t27c): refuse a Rust match block the parser kept as text; rewrite nine as switch #5964 (Closes t27c lowers a Rust match block to nothing, and typecheck accepts it #5949) base fix(t27c): refuse macro calls and typeless fields; rewrite eight Markdown specs (#5923) #5947's branch; fix(t27c): refuse a colon inside a field type and @as to a slice type; fix the six silent specs #5971 (Closes t27c: six silent misreads -- a colon inside a field type, and @as to a slice type #5968) is already squashed into it
• fix(gen-zig): keep _ = call(..); instead of deleting it (Closes #5984) #6003 (Closes gen-zig deletes every top-level _ = call(args); in a fn body (dead-store elimination treats _ as a store) #5984) gen-zig keeps _ = call(..);, base master
• gen-rust assert(c) -> assert!(c); (branch claude/gen-rust-assert-bang, PR to follow)

Follow-ups

Compiler (need cargo, so a Claude lane plus the lab, not a bee):

• gen-zig dead-store elimination also deletes y = call(..); when y is never read, call and all (the _ = half is fix(gen-zig): keep _ = call(..); instead of deleting it (Closes #5984) #6003)
• gen-verilog emits a raw _ = g(x); inside if / while bodies
• t27c seal --save writes the seal of a spec given by a temp path under that path's name

Spec source (a bee can do these with the prebuilt t27c):

• the 35 refused pairs tri misread --list names, one spec per task; the first is filed below
• two orphan governance_agent seal files named after a # comment

Master, not this stack (seen red on master 2e852cf82, listed so nobody blames the stack for them): duplicate-bodies (three bodies under local_branch, nested_tail, ordinary_if now copied twice and in no ledger), spec-guards, fpga-conformance.

Numbers at the start (lab run of f521a8a, the #5947 head)

• tri misread --list: 0 of 35 pairs silent.
• suite: RATCHET CLEAN, ledger 112 / 112.
• seal coverage: 1449 seals, 1325 hold, 124 known-broken.
• specs generate: 1251 specs, 1245 generate, 6 known-broken.

[gHashTag]

Progress, 2026-10-04 15:45Z

Edited 15:50Z: #5964 merged into #5947's branch while this was written.

Every number below is from a lab run linked next to it.

Lab (latest.json, PR #6097 head 5027a9f): the misread gate now reads both report formats (master's one-column tri misread and the stack's pairs refused silent table) and is left out of the verdict; zig 0.16.0 sits at /opt/zig for re-seals. A full run (build, suite, lean, seals, specs, misread) is about 3 min there; the same release build took 15 min 32 s on the workstation.

Stack

PR	head	lab	GitHub
#5947 (base master)	df32355 (now holds #5964)	pending	conflicts with master
#5964	merged into #5947's branch by the owner, 15:42:48Z, as df32355; its last head 821f358 was green	fpga-conformance red: elaboration ratchet 178 vs baseline 176 (ternary_isa 6 -> 7, uart 0 -> 1; run used iverilog 12.0, baseline 13.0). Cause not named yet.
#6003	merged as cdc02ad
#6103 (base master)	07557bc	green	master merged in on the lab: compiler.rs end-of-file hunk (both test modules), FROZEN_HASH, 3 seals re-sealed with zig
#6097 (base master)	5027a9f	green

tri misread

• master cedb826: 41 of 41 pairs silent; master's tool does not measure refusals (run).
• fix(t27c): refuse a Rust match block the parser kept as text; rewrite nine as switch #5964 821f358 (the stack's tool, which splits refused from silent): 0 of 35 silent, 35 refused (run).
• The two tools count pairs differently, so 41 and 35 are not the same population.

Master, not this stack: duplicate-bodies and spec-guards are red on master cedb826 as well as on #6097 and #6103; Scorecard analysis too.

Queen: 30 of 70 workers active. #6095 is in progress: branch queen-6095 exists, no commits on it yet. #6096 is in review on the board, though no pull request names it yet and the two orphan seals are still on master. Both are in skipSummary.claimed.

Steward: profile agent .claude/agents/t27c-steward.md (on #6097); scheduled task t27c-steward, every 2 h at :47, first round due 15:58Z. It runs while the desktop app is open.

Workstation load average: 212.

Next

1. Tell master's elaboration count from fix(t27c): refuse a Rust match block the parser kept as text; rewrite nine as switch #5964's: fpga-conformance on gen-rust: std macros keep their bang -- assert!, assert_eq!, assert_ne!, panic!, unreachable! (Closes #5987) #6103 (master plus a Rust-only change) is queued; 178 there too means the rise came from master.
2. Merge master into fix(t27c): refuse macro calls and typeless fields; rewrite eight Markdown specs (#5923) #5947 (now the only open compiler PR of the stack, and it carries fix(t27c): refuse a Rust match block the parser kept as text; rewrite nine as switch #5964's fpga-conformance result).
3. Read the bees' pull requests for sequence_hdc: replace the three map-typed fields with key/value slices (typecheck refuses the spec) #6095 and governance_agent: remove two orphan seal files named after a # comment #6096 against their Success Criteria.

[gHashTag]

Steward round 1, 2026-10-04 15:49-16:00Z

Moved: master cedb826 merged into #5947 as 655de092e; #5947 is MERGEABLE again. No new code was needed. The conflicts:

• cli/t27b/src/{lower,main}.rs: fix(t27c): refuse a Rust match block the parser kept as text; rewrite nine as switch #5964's squash (df32355) carried an older master snapshot of cli/t27b. No commit on the stack touches t27b, so master's version was taken; cli/t27b now matches master exactly.
• bootstrap/src/compiler.rs: an end-of-file append. Both of the branch's test modules stay ahead of master's tests_5984. The result is byte-identical to df32355's file, so FROZEN_HASH stays 57c987095b1e.
• wp18 selftest seal: the merged spec and compiler.rs are byte-identical to the branch's, so the branch's seal was kept. The lab's seal gates agree.

Stack

PR	head	lab	GitHub
#5947	655de09	green: suite RATCHET CLEAN, seals 1449/1325 hold/124 known-broken, specs 1259/1253/6	MERGEABLE; spec-guards red, also red on master cedb826 (see below); 13 checks pending
#6103	07557bc	green	MERGEABLE
#6097	5027a9f	green	MERGEABLE

tri misread: #5947 655de09 reports 0 of 35 silent, 35 refused. Master f244bd2 (green) reports 41 of 41 silent; master's tool has no refused column.

Master, not this stack: duplicate-bodies is red on master f244bd2 and cedb826. spec-guards is red on cedb826 and still pending on f244bd2.

Bee tasks: #6095 and #6096 are both claimed by the Queen. queen-6095 still points at old master 2e852cf with no commits, and #6096 has no branch or PR, so there is nothing to review yet. Queen: 16 of 70 workers active.

Workstation load average: 146 (1 min).

Next: wait for #5947's GitHub checks, then for the owner's merge. After that, open the first follow-up compiler lane (gen-zig y = call(..); dead-store), which is not started yet. The fpga-conformance question (178 vs 176) is still open; #6103's GitHub fpga-conformance is still pending.

[gHashTag]

Next step 1 of the 15:45Z comment, answered: fpga-conformance on #6103 (master cedb826 plus a change to Rust output only) reports the same elaboration ratchet as #5964 did, 178 (baseline 176), ternary_isa: 6 -> 7, uart: 0 -> 1 (job). The rise is not from this stack. It comes from master or from the runner's iverilog 12.0 against a baseline taken with 13.0; which of the two is still open, and it belongs in "Master, not this stack".