fix(lib): include any-to-any models in image-text-to-text filter

[cswaney]

Summary

• Adds an image-text-to-text entry to COMPATIBLE_PIPELINES that widens to any-to-any, mirroring the existing text-generation entry (which already lists both).
• The OCR batch launcher (and any other consumer of GET /api/models?image=image-text-to-text) now surfaces omni/multimodal models alongside VLMs.

Test plan

• just lint — clean
• just test — 960 passed, 7 skipped
• New API test asserts any-to-any models come back when filtering by image-text-to-text
• Verified against the running dev server: curl /api/models?image=image-text-to-text returned 3 any-to-any models alongside 57 image-text-to-text (before the fix, only the latter surfaced)

Closes #496

[claude]

Review

Small, well-scoped fix — the change and its test coverage line up cleanly with the stated problem.

Code quality

• The COMPATIBLE_PIPELINES["image-text-to-text"] addition is a direct, minimal fix and mirrors the existing text-generation pattern, as the PR description says.
• Nit (non-blocking): COMPATIBLE_PIPELINES now requires manually keeping any-to-any in sync across every pipeline key that should widen to it (text-generation and now image-text-to-text). If another omni-serving pipeline type shows up later (e.g. audio-text-to-text, video-text-to-text), it'll need the same manual entry. Not worth blocking on, but a small helper (e.g. {k: v + ["any-to-any"] for k, v in ...}) could remove the duplication risk going forward.

Bugs/correctness

• Logic looks correct: COMPATIBLE_PIPELINES.get(image, [image]) now returns ["image-text-to-text", "any-to-any"], so filtering by image-text-to-text surfaces the Qwen-Omni-style fixture model as intended. Verified by reading the query code at asgi.py:2125 / :2142.
• The fix is intentionally one-directional (filtering by any-to-any still won't surface image-text-to-text models), consistent with how text-generation already behaves — seems like a deliberate, acceptable asymmetry rather than an oversight.

Test coverage

• Good: the new test asserts against a concrete fixture model ID rather than just checking for presence of the tag, which guards against the widening including unintended models. Fixture addition looks consistent with neighboring entries.
• The PR description also mentions manual verification against the dev server (curl), which is a nice touch beyond automated tests.

Coverage badge

• lib/docs/assets/img/coverage.svg drops from 72% → 61%, which is a large swing for a 4-line functional change plus tests. Worth double-checking this reflects a real, pre-existing coverage baseline (e.g. badge was stale before this PR) rather than something going wrong in the regeneration step, since an 11-point drop could mask an unrelated coverage regression if it's real.

Security/performance

• No concerns — this only widens a read-only filter on an already-authenticated endpoint (/api/models), and doesn't expose any models that weren't already queryable through other filters (e.g. no filter, or text-generation).

Overall: looks good to merge pending a sanity check on the coverage badge number.