fix(#56): OpenCode context injection and MCP capture were both broken

[jsirish]

Summary

• Live-verified the OpenCode plugin for the first time since it shipped in v0.13.0 (issue OpenCode: live verification findings — context injection broken, MCP capture broken, skills confirmed not pluggable #56) and found two production bugs unit tests couldn't catch: session-start context injection never reached a real conversation (a decoy title-generation call consumed it first), and MCP tool calls were silently dropped (wrong id-naming assumption).
• Ports three session-start diagnostics the TypeScript plugin never had (version line, unconsumed-buffer sweep, gitignore nudge), wires post-compaction recovery, fixes a worktree-note symlink false-positive, bounds injected git status, and syncs the plugin's version (was 1.0.0 vs 0.13.0 everywhere else) with a 3-way version-agreement check.
• Closeslocal-ci parity gap: .opencode-plugin's npm ci/typecheck/test has no local-ci counterpart #49: local-ci now runs .opencode-plugin's own checks.
• README restructured so Claude Code / Codex CLI / OpenCode read as peer harnesses; docs/index.html updated to match.

Test plan

• npm test in .opencode-plugin — 89/89 passing (5 new tests covering the two production bugs + the ported behaviors)
• local-ci — all checks green, including the new opencode plugin custom check
• Both fixes confirmed live against a real OpenCode session (headless opencode run), not just unit tests — see OpenCode: live verification findings — context injection broken, MCP capture broken, skills confirmed not pluggable #56 for the full verification matrix
• TUI-only surfaces (compaction marker, first-run trust gate, skill picker) — in progress, tracked in OpenCode: live verification findings — context injection broken, MCP capture broken, skills confirmed not pluggable #56

Closes#49

🤖 Generated with Claude Code

https://claude.ai/code/session_018GvEcZ2V5qSX2KB1q2qLsu

[jsirish]

Ran the required `/review-pr` pass (4 parallel agents: code-reviewer, pr-test-analyzer, silent-failure-hunter, comment-analyzer). Findings and fixes, pushed in 291cbbc:

Critical — the P5 fix itself was buggy. Both code-reviewer and my own manual testing independently caught it: the worktree-sharing symlink fix canonicalized only root, not dataRoot, which flips the bug rather than fixing it (confirmed live: every plain macOS project under /var/folders//tmp falsely claimed worktree-sharing). Fixed at the source in lib.ts's computeDataRoot() to canonicalize on every return path, matching _lib.sh's tl_data_root() exactly. Re-verified live in both directions (plain project silent, real linked worktree still fires correctly) and added regression tests for both — pr-test-analyzer flagged this as a zero-coverage gap on a claimed fix.

Other fixes:

• Post-compaction recovery's path pointer was data-dir-relative instead of data-root-relative like every other injected path.
• NOISY_TOOLS was missing OpenCode's list/todowrite built-ins, which the broadened P0b default case would have started logging on every call.
• Several new failure paths (corrupt package.json, unreadable buffer, failed recovery read) swallowed errors instead of breadcrumbing to .capture-errors, the plugin's existing convention.
• Stale comments (removed mcp__ convention, old delete-on-first-read design), a naive path-prefix check, README's dangling session.idle cross-reference.

Also researched Jason's architectural question (should this shell out to hooks/*.sh instead of reimplementing in TS) via context7 + the upstream anomalyco/opencode repo. Filed #58 tracking anomalyco/opencode#12472 (native Claude Code hook compat — open, active, not shipped) as the trigger to revisit; not attempted now since no stable compat surface exists yet.

All 93 tests passing, local-ci clean.