feat(vision): cache vision-proxy image descriptions

[ChauThan]

Summary

Caches the text description produced by the vision proxy per image (SHA-256 of its base64 bytes), so images that were already described in earlier turns are reused without calling the vision model again. This saves Copilot quota and latency in multi-turn conversations where text-only models receive the same image attachments on every turn.

What changed

• New src/visionProxyCache.ts - an in-memory cache keyed by image content hash, capped at 200 entries with FIFO eviction (mirroring the reasoning-content cache).
• src/extension.ts - proxyVision() now:
  • describes only the message that contains a new image (instead of re-sending the whole conversation), and
  • reuses cached descriptions when all images in a message were already described - no vision-model request at all in that case, so selectChatModels()/sendRequest() are skipped lazily.
  • messages are flattened before the vision pass so the correct description is applied to the right apiMessage when one input message expands into several (e.g. tool results).
• src/test/visionProxy.test.ts - unit tests for the cache (stable keys, lookup/store, partial hits, reuse, FIFO eviction, clear).
• package-lock.json - version synced to 0.5.1.

Why

Without a cache, proxyVision() calls the vision model via model.sendRequest() on every turn to describe the same bytes again - wasting Copilot quota, adding latency, and returning a different description each time.

Validation

• npm run compile passes
• npm test passes - 168 tests (7 new for the cache)
• npm run lint:js passes

Issue

Closes #119

[ltmoerdani]

Nice one @ChauThan this is a clean fix. the cache mirrors the reasoning-content cache nicely, and the 7 tests cover the important cases (the partial-hit one especially).

a couple of things i want to double check before merging:

1. multi-image messages. you store the combined description under every image hash, so if one of those images shows up alone in a later message it gets a description that also mentions the other image. not a correctness issue since a partial hit still forces a re-describe, just confirming that's intended.

2. the proxy now describes only the message that has a new image instead of the whole conversation. that changes the behavior a bit, descriptions lose some conversation context. makes sense for the token savings, just wanted to flag it.

also, the lockfile sync to 0.5.1 fixes a mismatch that was sitting on main (package.json was already 0.5.1 but the lockfile was still 0.5.0), so that's a welcome side effect. the extra peer: true flags on a few dev deps look like npm version noise, harmless either way.

i'll run compile + the full test suite locally before merging. thanks for the PR.

[ChauThan]

@ltmoerdani
Thanks for the review!

1. Yes, that's intentional. I store the same description under every image hash. A partial hit returns nothing, so the message gets described again - no risk of wrong descriptions.

2. That was not intentional - I didn't think it through carefully. I'm thinking about adding a setting flag so users can choose whether to accept this trade-off or not. Thanks for pointing it out.

I'll do it if you agree.

[ltmoerdani]

@ChauThan thanks for the follow-up.

1. good, that matches what i read in the code. partial hit forces a re-describe, so no stale description gets reused.

2. sure, go ahead with the setting flag. a couple of guardrails so it stays small: keep the cache behavior as the default, gate the whole-conversation mode behind the flag, and add a short docs entry for it. ping me when it's ready and i'll review + test before merging.

[ChauThan]

@ltmoerdani done - added the setting flag as discussed.

• opencodego.visionProxyWholeConversation (default false) - off keeps the cache behavior (describe only the message with a new image and reuse cached descriptions), on describes the whole conversation so descriptions keep full context (still cached, more tokens).
• proxyVision() now takes describeWholeConversation; the whole-conversation branch sends one request over all messages and stores the combined description under every image hash (same no-partial-reuse rule as before).
• Docs: README settings table + a short section in the vision-proxy feature doc.
• compile / test (168 pass) / lint:js all green.

Ready for review + test whenever. Thanks!

[ltmoerdani]

@ChauThan nice one, this rounds it out well. i pulled the branch and ran the checks locally: compile passes, 168 tests pass (0 fail), lint:js and markdown lint on the feature doc are clean. the setting registration, README table, and the feature doc section all line up. merging now, thanks!