Regen normalization: memo fixed points, batch rustfmt per pass; seed build reads its compiled-crate count

[gunbai-bot]

Fast-regen layer 3 follow-up: two cost findings from the first //gunbc/instruments:regen-round-cost measurement (instrument-dispatch run 37293307189, on #13370's head). Both were scoped by sharp-raven-357.

1. rustfmt spawns: a bare-minimum-cost defect (DESIGN §6)

That round made 654 rustfmt spawns for 163 mirrors, roughly 58 s of mirror_write + digest. Two causes:

• Every distinct input paid its own fixed-point seek of at least two spawns: format, then confirm.
• The committed side (whose bytes ARE the previous normalized output) re-ran a seek the emitted side had just finished.

Fix, in required_regen_host:

• memo_normalized records a confirmed fixed point as itself. It's exact: the seek returns Y only once it has seen rustfmt(Y) == Y. A committed mirror equal to a normalized output then costs no spawn.
• normalize_population runs the same fixed-point seek over a whole population with one rustfmt invocation per pass. write_emitted_tree and tree_digest_for_basenames prefetch through it.
• It passes --config skip_children=true. Given a FILE, rustfmt follows mod declarations to sibling files, which stdin mode never does, and that refused a remote round with failed to resolve mod. On rustfmt 1.9.0-stable the option is accepted and the output is byte-equal to stdin. A rustfmt that ignored it would refuse loudly on the unresolved mod, never produce wrong bytes.
• Control: batched_normalization_is_byte_identical_to_the_single_seek uses fresh formatters with no shared memo. Every member's batched bytes must equal the single stdin seek's, and each output must be memoized as its own fixed point. The inputs include a member declaring a module with no sibling file. Run remotely and green. Unit tests are off CI (rust_unit_tests_off_the_merge_path), so this executed run is its evidence.

2. Seed build reported compiled_crates=0 while spending 655 s of CPU

The cause is the parse, not the build. The workflows set CARGO_TERM_COLOR=always, the round's cargo child inherits it, and cargo's Compiling lines arrive with escape codes, so the count read 0. seed_cargo_build now passes --color never.

The build itself DID compile, at 655 s CPU in 203 s wall. The next dispatch will say which crates. My suspicion is unconfirmed: the job builds -p v1-compiler --bin gunbc while the round builds --bin claim_executor without -p, so cargo may resolve features differently. Changing that waits for the evidence.

Declared

The three new host declarations get gunbc.regen_rustfmt_batch_seed_growth, enrolled in seed_growth_admission.

Evidence

• cargo clippy -p v1-compiler --all-targets -D warnings: clean (remote, on the merged-with-main tree).
• The control above: green (remote).
• The spawn count and wall-time effect will be measured by the next regen-round-cost dispatch. That runs after instrument-dispatch: admit //gunbc/instruments:regen-round-cost (priced regen round + one-mirror emit discriminator) #13370 lands, together with the two-mirror probe.

🤖 Generated with Claude Code