[codex] Add full manual repair workflows for completed runs

[CKwin26]

Summary

• add a companion autoresearchclaw CLI for manual paper-repair and research-repair workflows on completed AutoResearchClaw runs
• package the new repair CLI alongside the existing researchclaw entrypoint so users can initialize, publish, roll back, and relaunch repair workspaces from the same repo
• document the new repair workflows in the main README
• include the previously prepared repair hardening under the hood: ACP recovery/debug support, runtime asset/env forwarding, and stronger Stage 10 safety gates for reruns

Why

AutoResearchClaw already has in-pipeline rollback and auto-repair loops, but it did not ship a manual repair workflow for the cases humans actually hit in practice:

• a run completed, but the exported paper package still needs bounded cleanup
• a run completed, but the human reviewer wants a child rerun from Stage 9/10/12 without destroying the parent run
• a repaired rerun needs better ACP recovery, runtime asset injection, and stronger experiment safety checks to make that manual loop viable

This PR adds the missing operator-facing repair feature instead of only relying on internal rollback behavior.

What Changed

New manual repair CLI

Adds a new companion package and entrypoint:

• autoresearchclaw paper-repair-init
• autoresearchclaw paper-repair-apply
• autoresearchclaw paper-repair-rollback
• autoresearchclaw research-repair-init
• autoresearchclaw research-repair-run

These commands let a user:

• create an editable paper-repair workspace from a completed run
• publish repaired paper artifacts back into the source run with rollback snapshots
• create a research-repair workspace that preserves the parent run and prepares a child rerun from a later authoritative stage
• generate a child-run config, repair brief, launch script, and metadata without immediately launching a long job

Packaging and docs

• package the new autoresearchclaw module in the wheel
• add an autoresearchclaw console script alongside researchclaw
• document manual paper repair and research repair in the main README

Repair hardening included in the same branch

This branch also carries the repair/rerun hardening needed to make the feature practical:

• ACP reconnect/debug instrumentation and failed-prompt capture
• sandbox entrypoint arg/env forwarding so repaired experiments can receive runtime asset paths cleanly
• stronger Stage 10 validation for self-contained generated projects, placeholder/demo code, and ablation distinctness checks

Validation

• python3 -m compileall autoresearchclaw pyproject.toml
• python3 -m autoresearchclaw --help
• python3 -m autoresearchclaw paper-repair-init --run-dir <existing-run> --output-dir .tmp_repair_smoke/paper
• python3 -m autoresearchclaw research-repair-init --run-dir <existing-run> --output-dir .tmp_repair_smoke/research --config config.researchclaw.example.yaml --target-stage EXPERIMENT_DESIGN --reason smoke --feedback 'smoke feedback' --upstream-root .
• python3 -m autoresearchclaw research-repair-run --repair-json .tmp_repair_smoke/research/research-repair.json --skip-preflight
• plus the earlier repair-hardening checks on ACP/debug instrumentation, Stage 10 distinctness validation, and sandbox arg/env forwarding

Notes

• This is intentionally a companion CLI rather than a deep rewrite of the main researchclaw command surface.
• The manual repair workflows are additive: they complement existing gate rollback and auto-repair loops instead of replacing them.
• The branch includes the prior repair-hardening commit because the manual rerun workflow depends on those lower-level reliability improvements to be useful in practice.