fix(chat): ignore a stale voice recognizer's late onresult callback

[pedrofrxncx]

Follows #6494, which guarded onerror/onend in use-voice-input.ts against a stale SpeechRecognition instance firing async callbacks after a restart — but missed onresult.

Payoff: finalTranscriptRef/interimTranscriptRef are shared across recognizer instances. If a user stops/cancels and quickly restarts recording, the old (stale) recognizer can still fire a late onresult with a final transcript, which splices into the new session's transcript even though the old recognizer was supposed to be discarded — the exact same bug class #6494 fixed for onerror/onend, just on the one callback that was missed.

Fix: add the same recognitionRef.current !== recognition guard to onresult.

Regression test: use-voice-input.test.ts — starts recording, cancels, restarts (two instances), fires onresult on the stale first instance with a final transcript, and asserts the new session's transcript stays empty (previously it would have picked up the stale words).

Verify: bun test apps/web/src/hooks/use-voice-input.test.ts (6 pass).

Locally ran: bun test apps/web/src/hooks/use-voice-input.test.ts, bunx oxlint apps/web/src/hooks/use-voice-input.ts apps/web/src/hooks/use-voice-input.test.ts, bun run fmt. Full CI validates the rest.

---

Summary by cubic

Ignore late onresult from a stale SpeechRecognition instance after a restart to prevent stale text from splicing into the new session. Previously, a cancelled recognizer could append a final transcript to the new session; now the hook drops those events.

• Add recognitionRef.current !== recognition guard to onresult, matching existing onerror/onend guards.
• Add a regression test that starts, cancels, restarts, then fires the stale instance’s onresult and asserts the new session’s transcript stays empty.
• No API changes; only stale results are ignored.

^{Written for commit 3d58328. Summary will update on new commits.}