Plan for 0.3.0, and 0.4.0 as an indexing preview

[Seungpyo1007]

Adds docs/PLAN-0.3.md. Planning only — no code changes.

Why

The 45 commits on develop are all fixes: unsaved-edit loss, BOM and non-UTF-8 destruction, an approval prompt that stranded a parallel run, files invisible to the whole app, a watcher that truncated its own list. That makes a good changelog and a weak release — nothing in it is a new capability.

This document decides what 0.3.0 adds.

0.3.0 — three tracks

A. Providers (Gemini, local models) — the headline. OpenAICompatProvider is already config-driven, so Gemini is close to a config entry; the real work is local endpoints (Ollama, LM Studio, llama.cpp): lift the hardcoded URL into a setting, stop treating "no API key" as unconfigured, read /v1/models, disable the quota probe that means nothing locally, and say so when a small model cannot do tool calls instead of looking broken.

B. Editor-core visual effects (finishing Phase 2) — the first deliverable is a survey table of what exists versus what is missing. Listing candidates before that survey would mean writing down features that do not exist, and "finished" needs a definition first.

C. Extension shim gaps — WorkspaceEdit file create/delete/rename first. The first commit on this track is a README correction rather than a feature: the README says the remaining gaps "throw rather than returning an empty value", and decorations do not —

createTextEditorDecorationType: () => ({ dispose() {}, key: "sz-deco" }),   // vscodeShim.ts:557
setDecorations() { /* no-op */ },                                          // shimDoc.ts:125

That is the exact shape 0.2.0 existed to remove, still present, with the documentation asserting the opposite.

The cut line is written down too, so the release can end: A and C's README fix plus WorkspaceEdit are required, B's survey table is the floor, custom editors and the debug adapter API are decided on remaining capacity.

0.4.0 — codebase indexing, as a preview

Three fixes this round (depth cap, dot-directories, silent truncation) all came from one structure: the agent answers by walking the disk every time. Capping is unavoidable there, and with a cap "not searched" and "not present" keep arriving as the same answer.

The hard part is not building an index but keeping it honest. As confirmed this round, the watcher cannot report every name — the OS drops events first — so incremental update driven by watcher notifications is untrustworthy on its own. A stale index is worse than today's behaviour, which is why 0.4.0 ships it off by default, as a preview.