fix(dictation): stop deriving fallback target IDs from mutable placeholder text

[rosscado]

Summary

• getTargetElementId()'s fallback identifier for id-less dictation targets baked in the element's placeholder attribute — but UniversalDictationModule itself rewrites that placeholder mid-session (e.g. to "Recording...") to show dictation progress, so the key used to register a target's audio/refinement state drifts from the key later used to look it up.
• This makes contextual refinement (the second phase of dual-phase transcription) silently no-op on any field without a stable id — reproduced live on x.com/i/grok's "Ask anything" composer while scoping baseline universal-dictation support for Grok. The local test fixture never caught this because all its fields have stable ids.
• Fix: cache a per-element fallback ID in a WeakMap, generated once and independent of any later attribute mutation.

Test plan

• Added a fail-first regression test (DictationMachine-Refinement.spec.ts, "Target ID Stability (issue Dictation contextual refinement silently fails on any id-less field (target key includes mutable placeholder text) #507)") that mutates an id-less field's placeholder between segment capture and refinement; confirmed it reproduces the exact "No element found for pending refinement target" warning before the fix, and passes after.
• Updated an existing test that had encoded the old (buggy) placeholder-based key format as expected behavior.
• npm test green: tsc --noEmit, Jest (2/2), Vitest (205 files / 1795 tests, 1 pre-existing skip).

Fixes #507

🤖 Generated with Claude Code

https://claude.ai/code/session_01T4uuEsKVySbRV9XRQa2ebc

[rosscado]

Addressed the two review suggestions:

• Added a code comment on getTargetElementId() documenting the id-stability assumption (an element's id-presence must not change mid-session, or the same key-drift bug could recur).
• Noting here for the record: this fix also closes a pre-existing latent collision — two structurally-identical id-less elements (e.g. two blank <textarea>s with no class/name) previously mapped to the same fallback key under the old placeholder-based scheme; the counter suffix now gives each a distinct key. Side benefit, not the primary fix.