[slowdini]

What and why

Runner-dispatched Codex runs expose token usage but no native duration, which left all-Codex campaign timing summaries with zero duration samples. The universal runner already owns the subprocess lifecycle, so it can provide one cross-harness measurement without relying on transcript-specific timestamps.

This change measures and persists elapsed eval-agent subprocess time for every supported harness and carries that value into canonical timing.json artifacts.

Fixes #310.

Measurement contract

• Measure monotonic wall time only while the eval-agent harness subprocess is running.
• Sum the initial round and every native resume round.
• Exclude responder consultations, judge calls, queue time, and runner bookkeeping.
• Retain elapsed time for timed-out conversations; do not write a successful completion artifact after a failed round.

The measurement applies to Codex, Claude Code, Cline, OpenCode, and runner-ready BYOH descriptors without adding a harness-specific capability.

Artifact and compatibility behavior

• Add duration_ms to runner-owned conversation.json completion artifacts.
• Carry runner duration into timing.json while keeping canonical timing out of run.json.
• Replace the ambiguous shared source writer contract with independent token_source and duration_source fields, including the runner duration source.
• Continue reading historical records that contain only source; reserialization emits explicit per-metric provenance.
• Preserve native transcript duration as the fallback for historical and externally assembled runs.
• Leave historical Codex duration missing when no reliable timestamp exists; no runner time is fabricated or backfilled.
• Preserve an existing timing.json unless the operator requests overwrite behavior.

Before, a Codex timing artifact could not supply a duration sample:

{
  "total_tokens": 1234,
  "duration_ms": null,
  "source": "transcript"
}

Runner-driven ingestion writes independent provenance:

{
  "total_tokens": 1234,
  "duration_ms": 42891,
  "token_source": "transcript",
  "duration_source": "runner"
}

Aggregation checks source consistency independently for tokens and duration, so mixed provenance remains visible without conflating the two metrics.

Review notes

• The timing wire-compatibility logic lives in a focused core::timing module and remains re-exported through the existing core::types surface.
• Deterministic fixture delays cover every built-in harness plus BYOH, multi-turn accumulation, responder exclusion, timeout and failure paths, legacy deserialization, incomplete transcripts, overwrite behavior, and mixed-source aggregation.
• CLI help, shipped conversation and BYOH guides, harness notes, descriptor guidance, and JSON schemas document the boundary and fallback behavior.

Verification

• cargo fmt --check
• cargo clippy --all-targets -- -D warnings
• cargo test
• cargo run --quiet -- dispatch --help
• cargo run --quiet -- record-runs --help
• cargo run --quiet -- docs conversations
• cargo run --quiet -- docs byoh