[codex] Fix runner trace matrix and Hermes detection

[armand0e]

Summary

This patch release fixes the trace detection/conversion issues uncovered while testing the same two-turn prompt across every supported generation path.

• Bumps Teich from 0.1.3 to 0.1.4 in pyproject.toml, src/teich/__init__.py, and uv.lock.
• Recognizes Teich-exported Hermes external traces (external_session_meta with source: hermes-agent) as hermes, while still preserving the raw source metadata.
• Normalizes Codex conversion metadata to trace_type: codex and keeps Codex's native source separately as source: exec.
• Resumes Claude Code follow-up turns by native session id (--resume <sessionId>) instead of plain --continue, preventing Claude Code from injecting a synthetic "Continue from where you left off" turn and replaying the initial prompt.
• Chmods both external runner home directories and /workspace before sandbox snapshotting, fixing permission-denied failures after Hermes/Claude follow-up turns create files.
• Adds a generated Hermes example trace using the first examples/prompts.jsonl row, including its follow-up prompt.
• Adds focused converter/runner regressions for Hermes detection, Codex metadata, Claude follow-up resume, and workspace chmod behavior.

Root Cause

The live runner matrix exposed three issues:

1. Teich's Hermes runner exports traces as external_session_meta events with source: hermes-agent, but detection classified that shape as generic external_agent.
2. Claude Code follow-up turns used --continue, which can add an implicit continuation prompt and replay the original prompt before the real follow-up. Using the native session id with --resume preserves the intended prompt sequence.
3. Codex native session metadata uses source: exec to describe Codex CLI mode, not trace type. Conversion was surfacing that as metadata.trace_type, which made converted metadata disagree with detect_trace_type().

Validation

• Live runner matrix completed for all supported providers: chat, hermes, claude-code, pi, and codex using google/gemini-3.1-flash-lite via OpenRouter.
• Matrix validation confirmed each generated output converts to one row with turn_count=2 and includes both the initial prompt and follow-up prompt.
• uv run pytest -> 345 passed, 21 skipped.
• uv run ruff check .
• uv run python -m compileall -q src tests
• git diff --check
• uv build
• uv pip check

Notes

chat output is already structured training data, so detect_trace_type() correctly returns None for that file while conversion still accepts it as trace_type: chat.