fix: resolve all 21 sub-issues of friction-report tracker #245

[sriumcp]

Summary

External campaign-author friction report from running the
paper-memorytime-mirage campaign on nous against BLIS surfaced
21 distinct points of friction. This PR resolves the entire tracker
in a single coherent change, organized by theme.

• Spec-fidelity (F1, F2, F3, F4, F10, F13, F20): closes the gap left by HUMAN_DESIGN_GATE bypass under --auto-approve. campaign.locked_parameters ([F1] Add locked_parameters schema field to enforce campaign spec-fidelity in bundles #246) hard-fails bundle deviations regardless of --auto-approve. locked_workload ([F20] Schema: workload yaml deviations need locked_workload + workload_changes_from_canonical #265) does the same for workload yamls. Methodology prompt now declares the campaign > target-repo-docs hierarchy.
• Apparatus discipline (F7, F14, F16): methodology prompt edits cover the attribution-vs-totals distinction (with the BLIS runningBatch/RequestMap worked example), the unlocked-parameters audit, and the rehearsal-as-instrument positive pattern.
• Lifecycle / portability (F5, F11, F12, F19, F21): per-phase silence threshold ([F19] Per-phase silence threshold (DESIGN >= 600s, EXECUTE_ANALYZE lower); currently phase-agnostic #264/F19) closes the active stall where DESIGN's heavy reasoning trips an EXECUTE_ANALYZE-tuned watchdog. F21 introduces cumulative.patch + campaign.derived_from + nous lineage for cross-campaign code reuse.
• Reproducibility (F17, F18): reproducibility_metadata auto-captured at INIT (target repo commit, hardware-config sha, language versions, latency-config snapshots). nous package tarballs work_dir + reproduce.sh + Dockerfile + README.
• Hygiene (F6, F8, F9, F15): worktree_extras tracked-path warning, nous resume diagnostic, nous clean --orphaned, physical_realism_check schema.

The full per-F-entry → file map is in docs/friction-245-resolution.md. Every change is tagged in code with (#NNN / F<n>) so git blame + the issue tracker form a complete audit trail.

Why this PR is correct

A newcomer (or AI agent) navigating cold should be able to verify each F-entry's resolution in three steps:

1. Open docs/friction-245-resolution.md.
2. Find the F-entry's row; click through to the file/test the row references.
3. The code is tagged inline with (#NNN / F<n>) for cross-reference.

The architectural primitives are deliberately small and orthogonal:

• One validator function per F (in orchestrator/validate.py).
• One CLI subcommand per F when CLI-shaped (lineage, clean, package).
• One schema field per F with documentation in-place.

No refactors. Every behavior change is opt-in (legacy campaigns without locked_parameters keep working unchanged), and every schema addition is backward-compatible (oneOf'd into the existing fields where shape changed).

New modules

• orchestrator/reproducibility.py — F17 capture + per-iter snapshots
• orchestrator/lineage.py — F21 cumulative.patch + derived_from resolution
• orchestrator/plot_specs.py — F18 figure pipeline

New CLI surface

• nous stop --immediate — F5 event-boundary halt
• nous lineage <run_id> — F21 inheritance inspection
• nous clean --orphaned — F9 stale-worktree cleanup
• nous package <run_id> — F18 paper artifact tarball

New schema fields

Where	Field	Issue
campaign	locked_parameters	#246 / F1
campaign	locked_workload	#265 / F20
campaign	derived_from	#266 / F21
campaign	plot_specs	#263 / F18
campaign	reproducibility_metadata (auto-populated)	#262 / F17
campaign	sdk_timeouts.turn_silence_threshold_seconds (now accepts per-phase map)	#264 / F19
bundle	experiment_spec.physical_realism_check	#260 / F15
bundle	experiment_spec.unlocked_parameters_audit	#261 / F16
bundle	workload_changes_from_canonical	#265 / F20
bundle	experiment_spec.rehearsal_subset.depth_overrides	#248 / F3
bundle	timing_observations.recommended_turn_silence_threshold_seconds (per-phase map)	#264 / F19
state	reproducibility_metadata	#262 / F17

Test plan

• pytest tests/test_friction_245.py — 32 new tests covering F1, F3, F4, F11, F15, F17, F18, F19, F20, F21 (the F-entries with behavioral changes; F2/F7/F16 are prompt-text changes; F5/F6/F8/F9/F10/F12/F13/F14 are CLI/docs/runtime side effects).
• pytest tests/ — 1278 tests passing, 2 skipped, 0 regressions.
• python -c "import yaml, jsonschema; ..." — schema additions parse and validate against jsonschema.

Authoring discipline (for posterity)

The PR introduces docs/campaign-authoring-guide.md which captures:

• The "what to lock" inventory (avoiding the reactive failure mode that took five rounds to discover total_kv_blocks mattered in paper-memorytime-mirage).
• The pre-lock unit-check discipline (a closed-form sanity check before locking parameters).
• The rehearsal-as-instrument pattern (turning iter-1 into a diagnostic probe instead of a binary pass/fail).

CLAUDE.md and docs/architecture.md link to it.

Out of scope

uv.lock is untracked at session start (pre-existing). I deliberately did not commit it — that's a separate concern from #245 resolution.

🤖 Generated with Claude Code