feat(adapters): opencode dev/review synthesis parity (Phase 4)

[pbean]

Phase 4 of the opencode-http adapter plan (#163 → #164 → #165 → this): hookless dev/review roles now run the real bmad-dev-auto loop over HTTP/SSE.

What

• OpencodeDevAdapter(_DevSynthesisMixin, OpencodeHttpAdapter) — the Phase 2 synthesis mixin composed over the Phase 3 transport, paths= kwarg + _configure_dev_knobs(), mirroring GenericDevAdapter. Terminal-spec synthesis, stories mode, DW bundles, plan-halt, and the Hardening: reconcile seam is unreachable for stalled sessions — a finished-but-lost-Stop session cannot be recovered as completed #61 post-kill rescue all arrive via the mixin unchanged (no fork of the synthesis logic; devcontract.py and generic.py untouched).
• _probe_alive (the one seam the host must implement) = poll() on a per-task Popen stash populated in start_session and kept past kill() — the kill pops the _ServerSession registry before _post_kill_reconcile runs, and the retained handle answers liveness truthfully: never None (no probe that can hang), True when a force-kill-surviving process means the live-server invariant still applies.
• MRO note: the mixin's _result_json (spec synthesis) deliberately shadows the core adapter's result.json read-back — the dev skill writes no result.json, and a unit test pins that.
• _make_adapters: synthesizing hookless roles construct OpencodeDevAdapter (shared per (cfg, synthesizes) key); the temporary Phase 3 SystemExit guard is deleted, and cmd_validate's mirroring FAIL (added in feat(adapters): core opencode-http adapter over HTTP/SSE (Phase 3) #165's review round to match the guard) is removed with it.

Tests

FakeOpencode gains a spec-write mechanic (FAKE_OPENCODE_SPEC_PATH/_SPEC_TEXT, no-op when unset — existing scenarios byte-compatible). New coverage:

• E2E terminal-spec synthesis: scripted turn writes a status: done + ## Auto Run Result spec → synthesized dict matches devcontract (auto-dev, baseline_commit, story_key), usage captured, template-rendered prompt, clean teardown, and no FAKE_OPENCODE_RESULT_PATH set — proving dev never leans on result.json.
• E2E stories mode: BMAD_LOOP_SPEC_FOLDER id-keyed resolve with the mtime scan monkeypatched to raise.
• E2E post-kill rescue: spec written, idle never delivered (fake stays busy) → timeout under a live server → run()'s kill settles liveness → rescued completed with post_kill_reconciled: true.
• Units: dev knobs from policy, never-None probe tri-state, result.json-shadowing MRO pin.
• test_cli.py: guard tests flipped to their post-Phase-4 counterparts (dispatch + validate-passes).

All tests bind 127.0.0.1 and reuse the Phase 3 Windows-safe harness mechanics. CHANGELOG deferred to Phase 5 by plan design.

Verification

• Full suite: 2314 passed, 1 skipped
• trunk fmt (black) + full trunk check (no filter): clean

[pbean]

@CodeRabbit review

[coderabbitai]

Action performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.