Most turns. The parser’s farthest-error-wins rule cost the most. My first refusals sat in peek((KW_ASYNC …)) fail heads, and every one lost to a generic error. Moving them into named head rules didn’t help either. Only after several builds did I see that every fail has to stand after the input it consumes (async, then async(, then the option scan). The option scan then overshot to ) and needed a second rework. The interruption and the context restart that followed also cost turns.
Looked up. I had to find out how winnow’s fail turns into error: plus = help:, and that lowercase rules skip whitespace but a peek((...)) group does not. I had to find where immutable_item does its refusal and which checker file raises NK2103 (check_access.nika). I also had to learn that async(detached) needs fn main() throws (NK2605 otherwise), and that the specification baselines (EXPECTED.txt, COMPILES.txt) move when spec blocks start parsing. A few lines on each in the brief would have saved the search. The badge trap was already in builder-facts.
Checks. I ran the 15 async_parse tests with --both-settings on a scratch project, against the old binary and the new one. I also ran the whole it binary, specification::, clippy 1.97, cargo fmt and pre-push. The whole it run was more than needed: filters on overlap, spawn, task, async and specification would have been enough. Its one useful find was the two specification baselines. I skipped clippy 1.99.1 because of disk space. The full tests/compiler and tests/errors runs are left to the driver.
Errors in my report. It says Gate: green, but I never ran the gate; the field should say not run, driver’s. It also shows the project-test cache at 4.8 GB, while the limit is 2 GB. The cleanup step only deletes it when under 3 GB is free, so nothing was removed.
One change. Add a paragraph to builder-facts on refusing in the grammar. It should say that each fail stands after a consumed token, never inside a peek, because the furthest error wins and fail wins a tie. It should point to immutable_item as the model. It should also say that a newly parsed form changes the tests/specification baselines, so a builder regenerates them. That would have saved the most builds.
