fix: codex uses static bearer for inkwell MCP + retire legacy 'pcp' server (by Wren)

[conoremclaughlin]

Problem: ink -a lumen -b codex fails at TUI bootstrap with required MCP servers failed to initialize: inkwell: ... invalid_grant: Invalid refresh token.

Root cause: Codex independently manages its own OAuth credential for the inkwell MCP server (macOS Keychain) and refreshes it on startup. That refresh token has a 90-day server-side lifetime; once expired the server returns invalid_grant. In the codex TUI path MCP init is required, so codex aborts the whole session — even though ink already injects a valid Authorization bearer via env_http_headers. Codex was running a redundant OAuth dance anyway.

Fix:

1. Static bearer instead of OAuth (core fix). Switch both the CLI adapter (backends/codex.ts) and server runner (codex-runner.ts) from an Authorization env_http_header to codex bearer_token_env_var pointing at INK_ACCESS_TOKEN (raw token; codex prepends "Bearer "). This authenticates with the ink-owned token AND skips codex managed OAuth discovery/refresh, so the expired keychain credential is never consulted. (Same mechanism mcp.ts already uses when generating codex config.)
2. Retire the legacy pcp server name. Server was renamed pcp → inkwell. ink init/ink mcp sync already emit only inkwell, but injectSessionHeaders (shared/runner/mcp-config) and the gemini settings builder still fell back to a pcp server key when present, reviving stale config. Both now target inkwell exclusively (gemini-runner already did). No code path creates/feeds a pcp server block anymore.
3. Fix spawnSync sqlite3 ENOBUFS listing codex local sessions — default ~1MB buffer overflowed on large state_5.sqlite, silently dropping to the slower jsonl fallback and adding ~5s retry latency to every codex startup. Now 64MB maxBuffer.

Also removes the now-dead INK_AUTH_BEARER env var (mcp-config, codex spawn site in claude.ts).

Validation: reproduced with the expired keychain token still present — codex exec with the bearer override initializes inkwell MCP and replies normally (no invalid_grant, exit 0). Mirror-image confirmed: a same-URL duplicate server without the override still logs invalid_grant (non-fatal in exec, fatal in TUI), which is why the legacy pcp duplicate had to go.

• yarn build: clean (exit 0, 0 TS errors)
• adapters.test.ts + mcp-config.test.ts + spawn-reflect.test.ts: 47 passing

Operational note (not code): removed a stale [mcp_servers.pcp] block from the local ~/.codex/config.toml (same URL as inkwell, no auth injection, always OAuth-failed) so the interactive TUI boots clean.

🤖 Generated with Claude Code

— Wren