fix(core): keep partial Code Mode catalog updates proportionate - #57090

The following comment was made by an LLM, it may be inaccurate:

Posting numbers from the duplicate (#60129) as suggested, since they point the same way and go a bit further.

What we changed. CodeModeInstructions.update() falls back to render(current) whenever a shown listing changes while the catalog is partial. CodeModeCatalog.snapshot() now keeps the full inventory (path + rendered line) with a fingerprint, and select() derives the inline sample from it — so a rotated inline sample is no longer a catalog change. update() then diffs snapshots and emits only added/changed/removed listings and namespaces, restating everything only when the delta would be longer than a full render.

Measured on dev 726cf0963. Live opencode serve, 8 MCP servers (context7, enhanced-terminal, exa, gh_grep, graphify_hybrid, magpie-image, playwright, windows), disconnect all, then one connect per cycle ×6, one real step after each. Counting only the core/codemode notice text from the session DB (the same notices also carry core/mcp-guidance, split out):

packages/core codemode tests: 21 pass on dev, 31 pass with the patch; typecheck:tsc and bun run check clean.

Caveat on this scenario: on 726cf0963 an MCP disconnect does not shrink an already-created session's catalog while a connect does, so only the growth case is measured live.

Branch: codemode-delta-catalog on Neko-Yukari/opencode at 2208ed9e. Happy to fold this into this PR's shape if you prefer the smaller "short notice" approach over real deltas.

Numbers from folding this shape into the fuller delta design from #60129, so the two can be compared side by side.

What the merged behaviour is. A complete catalog always restates the added or changed listings, because search(...) is not offered there. A partial catalog keeps restating while the added and changed listings together stay under 1,000 characters — one listing costs less than the search step it saves — and past that it emits the namespace-level notice from this PR instead:

Tool listings changed in \bulk`. Re-run `search(...)` before relying on previous results.`

Measured on a 60-tool partial catalog (tools.bulk.toolN, signatures ~160 chars):

So the notice path costs roughly 1.5% of the block it replaces, and the single-listing path stays small enough that the model does not need an extra round trip.

Checks. packages/core test/codemode 34 pass, test/instructions 17 pass, typecheck:tsc clean. The new test covers both sides of the 1,000-character boundary, so a change that crosses into the notice path cannot silently flip back.

Where it lives. Neko-Yukari/opencode branch codemode-delta-catalog at 7c8bd7169, on top of 726cf0963, with the commit crediting you as co-author. The bulk of the branch is the snapshot/delta work; this PR's notice is the fallback for large churn inside a partial catalog.

Question. If maintainers would rather have the notice shape for every partial-catalog change and no restating at all, that is a two-line change in incremental() — the threshold is a single constant. Happy to reshape this branch that way, or to fold it into this PR if you would rather carry one diff.

@opencodejr flagging so triage sees this; if it belongs on #55529 instead, point me there and I will repost.

Thanks for the numbers. The threshold shape is the right call: restating under 1000 chars keeps the common single-change case cheap without an extra search round trip, and the notice covers bulk churn at about 1.5 percent of the block. I would like to fold it into this PR as one diff. Happy for you to fold your branch in, co-author credit stays of course.