perf(opencode): isolate per-proposal subprocess state to prevent SQLite WAL contention

[KE7]

Summary

• Fixes the SQLite WAL contention observed in the PR perf(evolution): atomic-worker architecture for parallel proposals (full GEPA parity) #34 E2E re-verify run where g1-s1 failed with "Failed to run the query 'PRAGMA journal_mode = WAL'" while g1-s2 succeeded — wasting one proposal slot
• Fix: set XDG_DATA_HOME to a per-candidate directory (<worktree>/.helix_opencode_state/) before invoking opencode run, giving each parallel worker its own isolated SQLite database
• 5 new unit tests; backend-exclusion test covers all 4 other backends (claude/codex/cursor/gemini)

Background

Depends on: #34 (perf/architecture-d-atomic-worker)

When num_parallel_proposals > 1 with backend = "opencode", helix spawns concurrent opencode run --format json subprocesses. All share the same global SQLite database:

• macOS: ~/Library/Application Support/opencode/opencode.db
• Linux: $XDG_DATA_HOME/opencode/opencode.db

Each subprocess issues PRAGMA journal_mode = WAL at startup. Concurrent requests on the same file produce the error seen in PR #34's re-verify.

Root cause archaeology: Confirmed via opencode debug paths that XDG_DATA_HOME changes the data directory (and therefore opencode.db path). Setting it per-candidate gives complete isolation.

Reference: PR #34 E2E re-verify report at /Users/ke/helix-arch-d-e2e-reverify-report.md:

g1-s1 status: Failed (SQLite WAL error — handled gracefully) ⚠️

OMAR comparison: OMAR uses OPENCODE_CONFIG_CONTENT env var for MCP config injection per-session, but does NOT set XDG_DATA_HOME. OMAR runs opencode in interactive TUI mode (tmux panes), not headless opencode run, so it doesn't hit the cold-start WAL contention. The XDG_DATA_HOME isolation is helix-specific and implemented directly.

Implementation

src/helix/mutator.py (~28 LOC including explanatory comment):

if backend == "opencode" and (sandbox is None or not sandbox.enabled):
    # Per-candidate SQLite isolation for concurrent opencode subprocesses.
    # [20-line comment explaining the WAL contention root cause]
    opencode_state_dir = Path(worktree_path) / ".helix_opencode_state"
    opencode_state_dir.mkdir(parents=True, exist_ok=True)
    backend_env["XDG_DATA_HOME"] = str(opencode_state_dir)

Also adds .helix_opencode_state/ to _ignore_helix_artifacts gitignore patterns.

E2E Validation Results

Both runs used the add_one off-by-one fixture with backend = "opencode", score_parser = "exitcode", frontier_type = "instance".

Config	Workers	Acceptance	SQLite WAL errors	Cost
n_proposals=2	2/2 ✅	100%	0 ✅	$0.00
n_proposals=3	3/3 ✅	100%	0 ✅	$0.00

Prior (PR #34 re-verify without this fix): n_proposals=2 → 1/2 (g1-s1 SQLite WAL failure).

Per-candidate isolation confirmed: each worktree has .helix_opencode_state/opencode/opencode.db.

Test Plan

• uv run pytest tests/unit -q → 871 passed (866 pre-existing + 5 new)
• uv run mypy --strict src/helix/ → Success: no issues found in 29 source files
• E2E n_proposals=2 → 2/2 success, no SQLite WAL errors
• E2E n_proposals=3 → 3/3 success, no SQLite WAL errors
• Reviewer verdict: APPROVE (opencode-harden-reviewer, /Users/ke/helix-opencode-harden-review.md)

🤖 Generated with Claude Code

[KE7]

Closing briefly to re-trigger CI after PR #34 merged into main.

[KE7]

Reopening; base auto-retargeted to main.