readme: agent trajectories section; reproduce: expected outputs

[mnkprs]

Why

The submission deliverables ask for representative trajectories for every agent, readable from the agent instructions to the final result, including the feedback that shaped the next step, retries and human checkpoints; and for a reproduction guide that says what output to expect. The trajectories existed but nothing pointed a reader to the representative ones, and the guide did not show expected output.

What changed

• README §2d "Agent trajectories": how a rendered trajectory reads (instructions → execution → result → verification), one representative file per agent and per kind of feedback (baseline, single agent with tools, judge + both probes, the two verifier-feedback corrective turns, tool errors, the strict-reviewer framing, real screenings from the skill), the three human checkpoints, and a "Where to go" row.
• README §2b: the in-session verdict is now the engine's exact printed output for examples/psf__requests-2317/ (same formatting as src/screen.ts) instead of a hand-shortened version.
• REPRODUCE.md: written-for-a-clean-machine line; expected terminal output for screen and the first lines of score on the committed results with the ±2-case expectation; pointer to §2d; stale §2b reference corrected to §2a.