Live / dirty-buffer indexing to eliminate index staleness (overlay + checkpoint)

[100yenadmin]

Prior art: SDL-MCP live-indexing (https://github.com/GlitterKill/sdl-mcp — see docs/feature-deep-dives/).

Gap in GitNexus: the index reflects committed state and lags HEAD (and lags unsaved edits entirely). Today's mitigation is "re-run analyze after a merge" + the freshness hook that notifies rather than auto-updating (to avoid LadybugDB write contention while MCP servers read). Result: agents query a graph that can be N commits / N edits behind reality, guarded only by the "trust d=1, verify against source" discipline.

Proposed feature — an optional live/overlay index layer:

• buffer/save pushes (debounced) → background tree-sitter parse → an overlay store where draft symbols shadow their durable counterparts → checkpoint into the durable index on idle (e.g. 15s).
• queries read the overlay first, falling back to durable.

Value: removes the staleness caveat that currently makes impact()/query() over recently-changed code unreliable. The small incremental overlay writes also sidestep the heavy-analyze write-contention problem that blocks auto-refresh today.

Pragmatic first step: even commit-granularity (incrementally index the working-tree diff vs the last-indexed commit) would close most of the gap and is simpler than keystroke-level live indexing.

[100yenadmin]

Source audit (read gitnexus/src/core/run-analyze.ts) — largely already covered; DOWNGRADING.

GitNexus already does incremental indexing of the dirty working tree: isGitWorkingTreeDirtyForAnalyze, incremental/subgraph-extract, shadow-candidates, and explicitly "even if HEAD matches lastCommit, the working tree may have uncommitted changes … fall through to incremental." Combined with the new post-merge incremental-analyze step (now in CLAUDE.md), practical staleness is largely handled.

The only residual gap vs SDL is the live push/checkpoint WITHOUT running analyze (sdl.buffer.push / sdl.buffer.checkpoint) — an IDE-integration nicety, not a token-saver. Low priority. The core capability already exists; closing-worthy unless we want editor-live overlays specifically.