batch(b): consolidate the 9 remaining reviewed PRs into one merge-queue entry

[noahgift]

Follows #2534 (batch A). Each merge-queue entry costs ~45 minutes and the queue is serial, so landing these nine individually is ~7 hours of CI; as one entry it is ~1.

Contents

PR	branch
#2478	fix/test-fixture-path-guard
#2484	fix/lockfile-sibling-guard
#2489	fix/tracked-ignored-scratch
#2493	feat/apr-data-alimentar
#2498	chore/delete-dead-duplicate-test-files
#2511	fix/apr-bin-prefers-fresh
#2515	chore/remove-whisper-apr
#2516	feat/expose-ptx-profile
#2527	fix/guard-batch

Seven merged clean. ci.yml conflicted in #2489 and #2527 — both the familiar case of two PRs appending independent guard-step blocks to the same job; both sides kept in each.

Two defects that exist ONLY in combination

#2493 and #2527 had each independently converted simular from a hand-rolled argv parser to clap derive. Both are correct on their own branch. Neither PR's own CI can see the collision, so merged separately they would have broken main.

1. Clashing public type names — the batch did not compile.

error[E0425]: cannot find type `Command` in module `simular::cli`
error[E0422]: cannot find struct `Args` in module `simular::cli`

#2493 dispatches to run_cli(simular::cli::Args { command }); #2527's API is Cli { command: Option<Commands> }. Took #2527's conversion — larger (129 vs 111 lines), ships the richer case table (ErrorKind::InvalidSubcommand, a SubcommandCase table), and it is the branch that owns the hand-rolled-parser ban this conversion exists to satisfy. Updated #2493's two call sites.

2. Duplicate help subcommand — clap debug_assert panic.

Command sim: command name `help` is duplicated

#2527's Commands carries an explicit Help variant. Standalone that is fine: its own Cli sets disable_help_subcommand = true. But #2493 embeds the Commands enum directly and never constructs Cli, so the attribute did not apply and clap generated its own help beside the explicit one. Set disable_help_subcommand = true at the embed site.

Worth stating plainly: clap's duplicate-name check is #[cfg(debug_assertions)]. A release build does not panic — it ships the ambiguity. This repo has already been bitten by that (duplicate short options in aprender-train-lora and trueno-zram). It surfaced here only because a debug test exercises the parser.

A third reported failure, every_lint_command_documents_its_flags, was collateral from the clap panic poisoning a thread join (parsing.rs:23 join: Any {..}). It passes in isolation and in the clean re-run — checked rather than assumed, since two failures could as easily have been two problems.

Verification — on the merged batch, rebased onto 7b2d0176a

cargo check --workspace --keep-going      0 errors
cargo test -p apr-cli --lib               7064 passed, 0 failed
cargo test -p aprender-simulate --lib     1853 passed, 0 failed
cargo fmt --check                         clean

check_readme_claims                       PASS
check_no_hand_rolled_parsers              PASS (ratcheted)
check_test_fixture_paths                  PASS
check_lockfile_no_registry_siblings       PASS
check_no_tracked_ignored_files            PASS (ratcheted)
check_apr_bin_resolution                  PASS
check_nextest_config_keys                 PASS
check_lockfile_current                    PASS

All eight guards these PRs introduce pass together — which is precisely what a batch can break and what individual PRs cannot reveal.

Cross-check that the counts stay coherent after the batch, using #2536's rebuilt README guard (now contract-based, 1s instead of a 14-minute build):

contract_count:    1778   (1772 + the six PCU contracts #2493 was missing)
cli_command_count:  111   (105 + #2493's newly-routed commands)

Both agree with the registry, so #2493's README claim is consistent rather than merely self-declared.

Note on #2515

Removing whisper-apr drops a crates.io dependency that was pulling a published copy of aprender itself into the lockfile, 36 minors stale. Measured against #2484's registry-sibling guard, that takes the count 7 → 1; the survivor (batuta-common via bashrs → aprender-compute-xtask) is the genuinely-transitive case that guard exists to catch. #2484 is shrink-only, so it reports an improvement rather than conflicting.