Bring media/json/tokenizer to 100% line/branch/function coverage: 97.06% → 100% branch - #1580

Approving. Baseline for every comparison below is current origin/main = aa18f265028eb3f6ba749e778b79b8dd8641bbf2 (Remove custom serialization infrastructure (Serializable trait) (#1579)).

This deletes a branch rather than covering one, so the burden is proving the arm dead — and dead on main, not only at the PR head. It is.

The reachability condition. parseMinusState's case '-' is entered only when the scan state is already '-' and the current JS token is '-'. State '-' is produced only by parseDefaultState on a '-' token, and (before this PR) by the removed arm itself. So the arm is reachable iff js/tokenizer can emit two consecutive - tokens.

Instrumentation on main, with a live negative control. Replacing the arm body with throw new Error(...) on main:

• full suite: 2797 pass, 0 fail — the throw never fires;
• driving the public tokenize() over 2,923,664 inputs (every ASCII string of length 1–3, plus every length-4 and length-5 string over - + 1 SP NL > < = * / . a " \ TAB): 0 throws.

Negative control, same harness, throw moved to parseMinusState's default: arm instead — the arm that now absorbs these inputs:

• full suite: 2789 pass, 8 fail;
• fuzz: 223,613 throws (first at input "-").

So the harness demonstrably reaches that switch; it is the case '-' arm specifically that is unreachable. Independently, scanning the same corpus for the token-level condition found 182,992 non-initial - tokens and zero adjacent -/- pairs — maximal munch turns -- into one '--' token (--- → '--', '-'), and - - gets a ws/nl token in between which routes through the default arm and resets the state to 'def'.

Two-sided equivalence. The removed arm and the default arm are not interchangeable (the removed one emitted one error token and stayed in state '-'; default emits an error plus mapToken(token) and returns to 'def'), so identical behaviour has to be shown, not assumed. Differentially running main's tokenize against the PR's over 4,720,084 inputs (the corpus above plus digits/e/E in the length-4/5 alphabet), comparing full token arrays and throw behaviour: 0 differences. Negative control on the differential (perturbing the PR side for inputs containing -): 660,901 differences, so the comparator is live.

Coverage claim measured, both sides. node --test --experimental-test-coverage --test-coverage-include='fjs/media/json/tokenizer/module.f.mjs' fjs/emergent_testing/all.test.mjs:

• main: 100.00 | 97.06 | 100.00 (line | branch | func)
• PR head: 100.00 | 100.00 | 100.00

The title's 97.06% → 100% is exactly right.

Battery. npx tsc --noEmit exit 0. npm run prepack from a freshly cleaned tree exit 0. node ./fjs/module.mjs t: 2797 pass, 0 fail — identical to main's 2797, as a dead-code removal should be. Broken-link sets diffed against main: identical. Changelog: changelog/unreleased/1580.md, one per-entry file, no headings, no links or PR numbers inside — correct shape, and its factual claim is the one verified above. No nanvm-lib/ or generator surface touched, so no Rust gates or round-trip needed.

One nit, not a blocker: the new comment says a run of - past the first "always falls through to the default arm below". For -- the state machine never enters the '-' state at all — the merged '--' token is handled by parseDefaultState's default. The emitted result is the same invalid token either way, so nothing behavioural rides on it.