jlens: resolve swap tokens exactly and suggest the closest token

[hijohnnylin]

Problem

When a jlens swap token was not in the vocabulary, the inference server fell back to the first token with the same trimmed text. It did not tell the user. So ants could swap in ants, ants\n or \tants, whichever had the lowest id. When no token matched, the user got a bare error with no hint. The fallback also scanned the full vocabulary (~150k entries) on every miss.

Change

Inference

• _resolve_steer_token_id takes exact matches only.
• On a miss, /v1/lens/prompt returns a 400 LensErrorResponse with {error, token, suggestedToken}.
• The suggestion is the longest vocab entry that is a prefix of the input. It tries the input as typed and with one leading space, and keeps a leading space that the user typed. The search is bounded: 3-54 µs on the Qwen3 tokenizer, including a 4,000-character input.
• openapi.json and lib/api/inference.d.ts are regenerated.

Webapp

• /api/lens/prompt forwards token / suggestedToken explicitly, and its swagger block documents them.
• jlens-stream.ts throws LensUnknownTokenError for that response.
• The steer panel shows "Not a single token. Closest: ␣antid [Apply]" below the swap input. Apply fills in the token and does not run the swap.

Compatibility

• New webapp with the old inference server: no change for users. Without the new fields, the plain error shows in the banner, as before.
• New inference server with the old webapp: the silent trim fallback stops, and a plain error shows with the suggestion in its text.
• Recommended order: deploy the webapp first.
• Public API: callers of /api/lens/prompt that relied on the trim fallback now get a 400. The 400 includes the suggestion.

Testing

• New apps/inference/tests/unit/test_lens_steer_token_resolve.py.
• Inference unit suite passes (643 tests, with HF_TOKEN), plus ruff, pyright and make openapi-check.
• Webapp npm run lint (eslint + tsc), format:check and npm test pass (154 tests).
• Checked suggestions and timing against the real Qwen3-8B tokenizer.
• Not tested end to end against a live inference server or in a browser.