session_diff stores full file contents via git show, causes multi-GB memory on resume #20990

Description

Snapshot.diffFull() in packages/opencode/src/snapshot/index.ts stores entire file contents in FileDiff.before and FileDiff.after via git show. These are persisted as JSON in ~/.local/share/opencode/storage/session_diff/<sessionID>.json and eagerly loaded on session resume.

In active repos with large files, these blobs grow to hundreds of MB per session. On our system:

593M  ses_3217a01c8ffeGbwxVoc2sfyYvb.json
593M  ses_321604e22ffeISyOc0LrkFJPUr.json
416M  ses_70361c737dbd27DXqU8FuEWBoA.json
...
44G   total in ~/.local/share/opencode/storage/session_diff/

Resuming a session with a 115MB diff file immediately spikes RSS to 4-10GB as JSON.parse() materializes all the strings. This is a major contributor to the memory issues tracked in #20695.

Root Cause

Two compounding issues:

1. diffFull() stores full file contents, not diffs. Lines 471-472 of snapshot/index.ts:

   git([...cfg, ...args(["show", `${from}:${file}`])]).pipe(Effect.map((item) => item.text)),
   git([...cfg, ...args(["show", `${to}:${file}`])]).pipe(Effect.map((item) => item.text)),

   For a 10MB JSON file, this stores 10MB in before and 10MB in after.

2. The diff covers the entire repo, not just the session's changes. diffFull() compares two git snapshots (from and to). In repos with concurrent activity (multiple agents, CI), the diff includes files modified by other sessions. A planning-only session that wrote 3 markdown files ends up with 22 files and 112MB of test fixtures it never touched.

   This is worse with jj (Jujutsu) co-located repos where the git working copy always has changes.

Evidence from heap snapshots

Heap snapshot analysis (uploaded to #20695) confirms FileDiff.before/after strings dominate the JS heap:

• Snapshot 1 (2.5GB RSS): 355MB in a single diffs Array(22923) — 51% of JS heap
• Snapshot 2 (11GB RSS): 7.64GB of strings, 96% growth from diff content
• TUI snapshot: 22 copies of a 20MB string retained by property:after

Suggested Fix

1. Don't store full file contents. Store unified diffs (git diff) or just file metadata (name, additions, deletions, status). The full content can be reconstructed from git on demand.
2. Scope diffs to session's own changes. Don't include files modified by other sessions/agents.
3. Don't eagerly load diffs on resume. Load lazily when the review UI actually needs them.
4. Cap content size. If before.length + after.length > limit, store only metadata.

Steps to reproduce

1. Start a session in a repo with large tracked files (>1MB)
2. Let other activity (commits, CI) happen in the repo
3. Close and resume the session
4. Observe RSS spike on resume proportional to session_diff file size

Environment

• OpenCode 1.3.13
• Ubuntu 22.04, 247GB RAM
• jj (Jujutsu) co-located repo with ~100 concurrent agent sessions
• 44GB total session_diff storage across all sessions