replace gemma image reading with moondream

[deathbyknowledge]

Summary

• replace the configurable Gemma image reader with the sole Workers AI model @cf/moondream/moondream3.1-9B-A2B, with no fallback path
• expand ai.image.read and img2txt across caption, query, OCR, point, and detect modes
• default img2txt to caption while keeping query prompts caller-owned
• add query reasoning/grounding, JSON/XML/Markdown/CSV structured output, schema checks, and mode-specific limits
• stream caption, query, and OCR through response bodies with cancellation and timeout cleanup
• preserve meaningful internal whitespace, remove old provider/model/input-format/prompt configuration, and simplify console settings
• update native command discovery, protocol/reference docs, and the shipped GSV Manual skill

Companion user/agent manual update: deathbyknowledge/gsv-manual#1

Validation

• gateway TypeScript check and full suite: 1325 passed, 5 skipped
• public SDK: 49 passed
• CLI/device Rust: formatting clean, 94 passed
• web: typecheck and production build pass; 277 tests pass; the unchanged ChatTranscript.test.ts bootstrap failure on DOMPurify().addHook remains reproducible on origin/main
• extension: check, 18 tests, and build pass
• assembler: 52 passed
• WhatsApp TypeScript and Discord/Telegram/test-adapter typechecks pass
• ripgit worker build passes; its configured Vitest run contains no test files
• focused bootstrap/config tests after the manual-skill update: 18 passed

[deathbyknowledge]

Live Workers AI testing found and fixed a response-shape mismatch in 54c08066:

• non-streamed env.AI.run() responses wrap Moondream fields under { result: ... }; the inference boundary now unwraps that envelope for caption, query/OCR, point, and detect
• streamed query/caption events carry cumulative text under chunk.answer / chunk.caption; the decoder now emits only each new suffix and ignores lifecycle/final aggregate events
• regression mocks now match the live binding shape

Validation: gateway TypeScript passes; 92 focused tests pass; full gateway suite passes (1325 passed, 5 skipped). A fresh CI run is in progress.

[deathbyknowledge]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a165e6b70d

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Enforce the submitted JSON Schema

When callers use standard JSON Schema constraints outside this small hand-coded subset—for example, { "type": "number", "minimum": 10 }—a Moondream result of 5 is accepted and returned as structured; object and array enum values are also incorrectly rejected because Object.is compares references rather than JSON values. This makes the advertised schema check unreliable for downstream automation. Use a compliant validator, or explicitly reject unsupported keywords and implement JSON-value equality for enums.

AGENTS.md reference: AGENTS.md:L43-L43

Useful? React with 👍 / 👎.