refactor(run-mode): remove the in-session dispatch mechanism (#119)

[slowdini]

Closes #119. Part of the Instant magic epic (#113).

Makes CLI dispatch the only mechanism end-to-end and deletes all in-session (DispatchMechanism::InSession / RunMode::Interactive) code. Claude Code now defaults to and runs via its claude -p stream-json path, exactly like Codex's codex exec. hybrid and headless are the only run modes (kept as accepted --run-mode values pending the vocabulary collapse in #120).

Net −2237 lines across 40 files — mostly deletion.

What changed

Commit 1 — delete the in-session-only subcommands. switch-condition and reset-batch were referenced only by in-session runbook generation and their own tests.

Commit 2 — remove the mechanism (atomic).

• DispatchMechanism enum, RunMode::Interactive, RunMode::mechanism() removed; default_for(ClaudeCode) → hybrid; resolve_run_mode ClaudeCode → [hybrid, headless]; HarnessRunCapabilities.mechanism dropped.
• Single shared env/ layout gone — every run stages one env-<group>-<condition>/ per (group, condition).
• HarnessAdapter: parse_cli_events* now required; runbook_template / parse_transcript* removed. Codex parses its events file directly; OpenCode errors until ingest is wired. Runbook always uses the shared headless template.
• Deleted in-session transcript code (subagent resolution by agent_description, resolve_subagents_dir*, SubagentMeta/SubagentEntry), the --subagents-dir / --session-id flags, and CLAUDE_CODE_SESSION_ID auto-resolution. Shared extract_invocations/read_records/TranscriptSummary kept (the stream-json parser reuses them).
• Pipeline record-runs/fill-transcripts read each task's outputs/<harness>-events.jsonl only.

Per the agreed scope: dead flags removed outright; only broken prose refs fixed (manifest, --help, stray-writes hint). The run-mode vocabulary/docs rewrite — including docs/harness-parity.md, which still describes the removed dispatch-mechanism API — is left for #120 / #121.

Before / after

A bare eval-magic run --harness claude-code (the default):

	Before (interactive)	After (hybrid)
Env layout	single iteration-N/env/	iteration-N/env-<group>-<condition>/ per pair
RUNBOOK	inside env/, agent-followed	iteration-N/RUNBOOK.md, human-followed CLI recipe
Dispatch	in-session Task subagents	claude -p stream-json per task
Transcript source	subagents dir via CLAUDE_CODE_SESSION_ID	each task's outputs/claude-events.jsonl

Verification

• cargo test — 468 pass, 0 fail (346 lib + 58 cli + 64 run + doc)
• cargo fmt --check — clean
• cargo clippy --all-targets -- -D warnings — clean
• grep over src/ — no InSession / DispatchMechanism / RunMode::Interactive / deleted-flag references
• e2e smoke (Claude Code claude -p): run → simulated dispatch (claude-events.jsonl) → ingest (assembled run.json + timing.json with source: transcript from the stream-json result event) → finalize (wrote benchmark.json)

Follow-ups (out of scope here)

• Likely supersedes Codex fully interactive support #34 (Codex fully-interactive support).
• Unblocks refactor(run-mode): collapse the run-mode vocabulary and rewrite the mode docs #120 (run-mode vocabulary + mode docs) and epic: data-driven harness descriptors (instant magic, phase 2) #122 (data-driven harness descriptors).
• docs/harness-parity.md rewrite owed by docs(harness): define the minimal harness interface + progressive-enhancement tiers #121.
• agent_description is now written to dispatch.json but no longer read (kept to avoid schema churn here); a candidate for later cleanup.

🤖 Generated with Claude Code