[BUG] User profile learning fragments profiles, never drains backlog, and can flip language

[ZeR020]

Surfaced by @ovizii's report in #55 (1,524 prompts stored, 1,484 awaiting, 40 analyzed → four separate anonymous profiles, one in the wrong language).

Bug 1 — Profile fragmentation when git email is unresolvable.
performUserProfileLearning falls back to userId = anonymous-${Date.now()}-${random} when tags.user.userEmail is empty (userEmailOverride unset + no git config user.email — common for non-coding use). A new random identity per run means getActiveProfile() never matches, so every 10-prompt batch creates a brand-new anonymous profile instead of updating one. 40 analyzed prompts = 4 batches = 4 profiles, exactly as reported.

Bug 2 — Analysis backlog never drains.
Each session-idle run analyzes exactly one batch of userProfileAnalysisInterval (10) prompts. A backlog of N prompts needs N/10 idle cycles; 1,524 stored prompts ≈ 148 idles. Learning permanently lags behind.

Bug 3 — Profile language can flip between batches.
The system prompt asks for "the same language as the user's prompts", but a mixed-language batch lets the model pick a different language per batch. Output language should be pinned (dominant language detection; existing profile's language when updating).

Fix direction: stable fallback identity (reuse latest active profile's userId, else constant anonymous), a per-idle drain loop bounded by a new userProfileMaxBatchesPerIdle config, and explicit language pinning via the existing language-detector service.

[ZeR020]

Fixed on main in 0becde9 — all three defects, verified by new regression tests (812 passing):

1. Stable identity. One shared resolveProfileUserId now backs every profile lookup (learning, dashboard get/refresh, the profile tool): your git email when available → otherwise the most recent active profile → otherwise a single stable anonymous identity. The per-run random anonymous-<timestamp> minting is gone, so no more one-profile-per-batch. Bonus: the profile tool no longer refuses to save a preference when no email is resolvable.

2. Backlog drains. Learning now loops FIFO batches per idle until fewer than userProfileAnalysisInterval prompts remain, capped by a new userProfileMaxBatchesPerIdle setting (default 5, i.e. 50 prompts/idle out of the box — raise it to catch up faster). One summary toast per run instead of one per batch.

3. Language pinned. Updates always keep the existing profile's language; new profiles are written in the dominant language detected across the whole batch (majority vote per prompt, reusing the existing language detector).

Two deliberate non-goals, so they're on record: pre-existing duplicate anonymous profiles aren't merged (the resolver self-heals onto the newest one going forward — for @ovizii this is moot since you already deleted the profile DBs); and if you later set userEmailOverride/git email after learning anonymously, a fresh email-keyed profile starts (the old anonymous one stays but is no longer selected). Ships with the next release; changelog updated.