core: session never drains when Snapshot.capture() stalls on huge untracked tree — git add hangs, execution claim never released

Description

Summary

Sending a prompt in a project with a very large untracked (and not gitignored)
file tree leaves the session stuck forever with no response. SessionStep.attempt
runs Snapshot.capture() before llm.stream(), and the capture's
git add --all --pathspec-from-file=- burns 100% CPU in pathspec matching for
10+ minutes (167k untracked files). No model call is ever made, and the
time_suspended execution claim is never released, so the session never drains.

Expected Behavior

The step either completes the snapshot in reasonable time, skips/times out a
pathological capture and still calls the model, or surfaces a snapshot error —
and the execution claim is released so the session stays usable.

Actual Behavior

• The runner claims the session (session_v2.time_suspended set 64ms after
  creation) and blocks in Snapshot.capture():
  packages/core/src/session/runner/step.ts:72 runs before llm.stream()
  at step.ts:102, so the provider is never contacted.
• Git.index.refresh (packages/core/src/git.ts:381) enumerates all untracked
  files via ls-files --others and feeds ~167k paths to a single
  git add --all --sparse --pathspec-from-file=- --pathspec-file-nul
  in the snapshot repo (~/.local/share/opencode/snapshot/<project>/<hash>/).
• That git add (child of the server process) ran 11+ minutes at ~99% CPU;
  sample shows 100% of time in read_directory_recursive →
  treat_path → match_pathspec_with_flags → strncmp — i.e. pathspec
  matching cost, not I/O. It also holds index.lock in the snapshot repo.
• time_suspended is never cleared (time_updated frozen), no
  Failed to drain Session error is logged, the UI just shows nothing.
  Side symptoms: a transient GET /api/session/<id> 404 (request raced
  session-row insert by ~74ms) and SSE reconnect churn, both harmless.

Additional Context

• Frequency: deterministic given the tree shape; every step start re-captures,
  so restarting the server re-triggers it (startup recovery resumes the
  claimed session).
• Workaround: kill the stuck git add, add the directory to the project's
  .gitignore, then restart the server / resend.
• Related (same feature area, different mechanisms, not duplicates):
  #46138 (Snapshot.diffFull serializes full patch text into message rows),
  #47854 (EventV2 duplicates large snapshot payloads),
  #47022 (~72GB DB growth),
  #47605 (sessions stuck busy via model-response timeout),
  #47988 (cancelled task blocks resume/amend).
• Possible fix directions (suggestion only): cap/skip capture when the
  untracked candidate set exceeds a threshold (log + continue with
  snapshot: undefined, which capture() already supports); add a timeout to
  Git.tree.capture; avoid feeding hundreds of thousands of explicit
  pathspecs to one git add (e.g. rely on directory-level adds or
  status --porcelain cone/sparse scoping).

Plugins

nothing globally; project-local .opencode/plugins/use-vp.ts present

OpenCode version

local (dev install, bun run src/index.ts serve) checkout of v2.0.2

Steps to reproduce

1. Have a git project with a huge untracked, not-ignored directory
   (in my case shadcn_ui_components_2026/: 167,696 files, 685MB;
   167,697 untracked vs 502 tracked; the project .gitignore did not cover it).
2. Open the web UI, create/select a session in that project, send any prompt.
3. Observe: user message is persisted (session_message has the user part),
   but no assistant message ever appears and no model request is logged.

Screenshot and/or share link

No response

Operating System

Darwin 24.6.0 arm64

Terminal

web UI in browser (TERM_PROGRAM not set in shell env). Shell: /bin/zsh