json_rpc: look up handlers with at, not bracket indexing - #1514

Approving. I treated this as a behaviour change rather than a refactor and reproduced the before/after directly instead of reasoning about it.

The bug is real, and worse than "unsafe indexing". Same probe (dispatch({ping}), request {jsonrpc:'2.0', method:<name>, id:4}), run against origin/main (37db36c) and against this head:

So main is not merely permissive: five of the six prototype names throw out of a function documented as pure and total, and toString emits a response that violates the errorResponse schema (error is the string "o" — "[object Undefined]" destructured character-wise). Every one of these is reachable from untrusted wire data. The todo file this PR deletes described exactly the two shapes I reproduced.

Not over-corrected either: an own toString handler still dispatches ({result:"own"}) on both main and this head, and a notification (id absent) still short-circuits to null before the lookup on both. at uses getOwnPropertyDescriptor + fromUndefined, so an own property whose value is undefined still yields -32601, matching the old === undefined test — the only behavioural delta is the inherited-name one.

The new proofs are not vacuous. I mutated the fix back (const handler = handlers[method] ?? null, verified the edit landed) and re-ran node ./fjs/module.mjs t: 2 failures — inheritedToString ("assertion failed") and inheritedConstructor ("TypeError: handler is not a function..."), i.e. the two distinct failure shapes each get their own guard. Restored, 0 failures. ownPrototypeName passes under the mutation too, but it is a regression guard for the opposite direction, which is the right thing for it to be.

Worth saying explicitly since coverage is now measurable: coverage could not have caught this. npm run cov reports fjs/protocol/json_rpc/module.f.mjs at 100.00 / 100.00 / 100.00 on main as well — the buggy line was already fully covered. The behavioural cases are the only thing that could have found it, which is an argument for this PR's shape rather than against it.

Verified at 3dd34d0, both trees clean before testing:

• npx tsc --noEmit — exit 0 (also 0 on main).
• npm run prepack from a freshly cleaned tree — exit 0 (both passes). Also 0 on main.
• node ./fjs/module.mjs t — 2498 pass / 0 fail vs 2495 / 0 on main. +3 is exactly the three added cases, nothing else moved.
• npm run cov — tests 2434, pass 2434, fail 0 (main: 2431), all files 99.93 / 98.07 / 99.73 — byte-identical totals to main.
• Dual-axis public surface (prepack in both trees, then extract.mjs / consts.mjs, outputs 45.9 KB / 64.7 KB so non-empty): zero differences in exported type aliases and zero in exported const signatures. No new export type — so no §6.2 _-prefix question arises — and dispatch's emitted signature is unchanged, no widening to any. Dropping the hand-written /** @type {Handler | undefined} */ cast costs nothing at the boundary; Handler is still exported from types.ts via Handlers.
• linkcheck.mjs — broken-link sets identical to main, not just counts. Deleting todo/dispatch-at-lookup.md strands nothing: grep -rn dispatch-at-lookup --include='*.md' finds no remaining reference (the cross-link ran the other way, from the deleted file).
• Rust gates skipped — nanvm-lib/ untouched.

Two smaller things, both correct:

• The ## 0.44.0 heading restore is a genuine fix, not a rewrite of a released section. #1513 added the line as bare text 0.44.0 under ## Unreleased; every other release in the file is a ## heading (## 0.43.1, ## 0.43.0, …). Restoring it is what makes the new Unreleased entry land in the right section.
• CHANGELOG entry links only /pull/1514, no issue or todo/ link, and sits under ## Unreleased. No **BREAKING CHANGES:** prefix, which I agree with: the exported surface is provably identical and the change only replaces a throw / malformed response with the documented -32601.

Nothing to fix. Comment in module.f.mjs naming the hazard matches the three existing at dispatch sites (effects/match, cli/dispatch, cas/evo), so json_rpc is no longer the odd one out.