feat(opencode): transcript ingest via the opencode-events parser

[slowdini]

What this does

Two scoped commits advancing OpenCode from staging-only toward full enhancement parity (model in #113):

1. fix(opencode) — closes Opencode skill name #65. OpenCode's skill tool lists frontmatter names, and staging rewrites the staged skill's name: to the sanitized slug — so the <available_skills> block in dispatch prompts now advertises the slug (as Codex already does) instead of the natural name that no longer resolves.
2. feat(opencode) — transcript ingest. Wires the opencode-events parser capability: opencode run --format json envelopes are self-contained, but tool args nest at part.state.input and timestamps are epoch-ms numbers — beyond the extract tier's five primitives, exactly the case the docs reserve for a code capability. The parser normalizes tool_use parts (name at part.tool, args = state.input, result = state.output/state.error), sums step_finish.part.tokens (cache reads excluded, matching codex accounting), recovers the final message from the last text part, and measures duration from envelope timestamps.

With this, OpenCode flips Transcript ingest to ✅ in the README support table. Remaining columns are filed: dispatch recipes + model flag (#153), shadow preflight (#154), write guard (#155) — each scoped with verified CLI/source evidence.

Details

• harnesses/opencode.toml declares [transcript] + the verified [tools] vocabulary (bash is the shell tool id; edit/write take filePath; apply_patch takes patchText — ids and arg shapes checked against packages/opencode/src/tool/ in the opencode repo).
• The __skill_invoked meta-check generalizes from hardcoded Skill/skill to descriptor-declared skill_tool/skill_arg (defaults keep claude-code.toml untouched; OpenCode declares skill/name — its native skill tool is a deterministic invocation event, so surfaces_skill_invocation = true and the meta-check grades from the transcript instead of the LLM-judge fallback).
• Shared write-boundary policy reads OpenCode's camelCase spellings (filePath in path_arg, patchText in apply_patch_paths) — additive, no effect on claude/codex classification.
• The transcript-less preflight-warning unit tests move to the byoh integration tests (all built-ins are now transcript-wired); opencode next-steps/runbook goldens re-blessed for the ingest-wired handoff text.

Schema change: harness-descriptor.schema.json gains opencode-events in the parser enum and optional skill_tool/skill_arg strings on [transcript]. No CLI change.

Before / after (generated next-steps)

- Next: ... dispatch each task with `opencode run`. OpenCode transcript ingest is not yet wired, so
  assemble each task's run.json/timing.json manually ...
+ Next: ... dispatch each task with `opencode run --format json`, saving stdout as the task's
  outputs/opencode-events.jsonl ... Then run `ingest ... --harness opencode`.

Verification

• cargo test — 657 passed, 0 failed (478 lib / 99 cli / 80 run), including a new end-to-end test: staged run → simulated opencode-events.jsonl → ingest yields parsed invocations, 125-token / 5s backfill, transcript final-message fallback, and a code-graded __skill_invoked (passed: true, grader: transcript_check).
• cargo clippy --all-targets -- -D warnings — clean. cargo fmt --check — clean.
• Both commits individually green (commit 1 verified at its own tree state).
• Docs updated: README support table + intro, docs/opencode-notes.md (code map / what's wired / next enhancements), docs/byoh.md, docs/progressive-enhancements.md, harnesses/template.toml, harness-descriptor PR template.

Event shapes verified against the opencode SDK types (ToolPart/StepFinishPart) and run.ts in anomalyco/opencode v1.18.3, per the don't-guess guardrail; a live opencode run capture is part of #153's acceptance (dispatch recipes) and will re-confirm the fixtures then.