Reconcile four parallel branches before adding more

[di-omics]

Four PRs are open against main, written by sessions that could not see each other. They are all additive -- no branch deletes another's work -- but they overlap enough that merge order matters, and two of them duplicate.

Shared files (mechanical conflicts, all four)

README.md, autonomous_lab/__init__.py, autonomous_lab/cli.py. Every branch adds exports and a subcommand. Conflicts are real but boring: keep both sides.

Real overlaps

#4 ⊃ #5	05476ed (workcell scheduler) sits on both. It landed on #4 by accident when a concurrent session switched branches mid-write, then was cherry-picked to a clean branch as #5. Same ~560 lines in two PRs.
#3 vs #5	orchestrator.py (2,026 lines, runtime policy-gated execution) and orchestrate.py (300 lines, static planning). One letter apart, adjacent concerns. They are genuinely complementary -- one executes, one answers "what serializes and what should I buy" -- but nobody reading the repo will guess that from the names.
#3 vs main	#3 also rewrites intelligence.py, provenance.py, throughput.py. Largest blast radius of the four.

Two other branches exist without PRs -- codex/evidence-learning-kernel and laboratory-intelligence-layer -- both based on cdfed4e, three commits stale. They add samples.py, evidence.py, gates.py, learning.py, criteria.py, acceptance.py, permission.py, control.py, record.py. Diff them against their own merge-base, not against main -- against main they appear to delete qc.py, vision.py, recovery.py, and throughput.py, which they never touched. Those files simply did not exist when the branches were cut.

Duplicated concepts across all branches

• Sample tracking: lineage.py (main) · samples.py (×2, different implementations) · sample logic inside orchestrator.py
• QC / criteria: qc.py (main) · gates.py · criteria.py + acceptance.py
• Authority: authority.py (Printed fixtures, and the path from video capture to a trusted arm policy #6) · permission.py + control.py
• Provenance: provenance.py (main) · record.py

Suggested order -- ascending blast radius

1. Compose vision with gates, and cover what a workcell needs to keep running #5 -- smallest, purely additive, CI green.
2. Printed fixtures, and the path from video capture to a trusted arm policy #6 -- additive, no edits to existing modules. ~28 audit majors documented in a PR comment are worth a pass first.
3. Check the ambition against the published record, not just against this lab #4 -- after dropping 05476ed: git rebase --onto main fd4ffcb~1 agent/standalone-readme. CI currently red.
4. Schedule more than one plate, and price a plate-moving arm honestly #3 -- last. It edits three existing modules, so it should rebase onto whatever the others leave behind rather than the reverse. Resolve the orchestrate/orchestrator naming here; renaming the planner to planner.py would make the split legible.

Then decide what to keep from the two stale branches before anything else lands, because a fifth sample tracker helps nobody.

The general fix

Branch from current main, and check git branch -avv before starting. Most of this cost came from branches cut at different points and from two sessions writing the same concept an hour apart.

[di-omics]

PR numbers changed; the branches did not

The PR and issue history was reset at some point overnight and main was rewritten with new SHAs. No code was lost. All eight branches survive on the remote with their commits intact, and every one of them still shares a merge-base with the rewritten main, so they remain mergeable.

The PRs were recreated. New numbering:

was	now	branch
#5	#3	feat/workcell-scheduling
#6	#4	feat/evidence-authority-cadence
#8	#5	feat/coverage-durability-teaching-feedback
#9	#6	feat/printed-fixtures (stacked on #5)
new	#2	feat/information-per-dollar

What did not survive is the review context: the audit findings recorded as PR comments, and the note on the README branch about its duplicated scheduler commit. The branch state is unaffected either way.

Revised merge order

Unchanged in substance, renumbered:

1. Schedule more than one plate, and price a plate-moving arm honestly #3 -- smallest, additive.
2. Compose vision with gates, and cover what a workcell needs to keep running #5 -- four new modules, no edits to existing ones.
3. Printed fixtures, and the path from video capture to a trusted arm policy #6 -- retarget its base to main once Compose vision with gates, and cover what a workcell needs to keep running #5 lands: gh pr edit 6 --base main.
4. Name the number every other layer is instrumental to #2 -- independent of all of the above; imports only from main. Can go any time.
5. Check the ambition against the published record, not just against this lab #4 -- additive.
6. agent/closed-loop-intelligence and agent/standalone-readme last, since the first edits three existing modules.

Still open from the original post

The duplication inventory has not changed: three sample-tracking implementations, three QC/criteria layers, two authority models across the unmerged branches. #2, #5 and #6 add no new duplicates -- coverage and information compose the existing layers rather than reimplementing them, which is the pattern the others should converge on.

Two branches remain based on a stale commit. Diff them against their own merge-base rather than against main, or they appear to delete files they never touched.

[di-omics]

Update: all five PRs rebased onto current main, all mergeable, all green

main moved (bundled evidence manifest, doctor needs no checkout) and left #3, #4 and #5 CONFLICTING. All are now rebased and passing.

PR	Base	Tests	Checks
#2 information per dollar, world model	main	214	green
#3 workcell scheduling	main	160	green
#4 evidence, authority, cadence	main	292	green
#5 coverage, durability, teaching, feedback	main	351	green
#6 printed fixtures, capture to policy	#5	477	green

One thing worth knowing about #6

Rebasing it surfaced a defect I introduced. An earlier commit on that branch used git add -A and swept in a concurrent agent's in-flight rename of plr-tested to "run-card checkout" across registry.py, model.py, ledger.py, workcell.py, throughput.py, protocols.py, cli.py and test_lab.py. None of it belonged to the printed-fixture work, and main has since gone a different way on the same question.

Every one of those files is now reverted to main, keeping only what the branch actually contributes. The lesson generalises: in a repo with concurrent agents, git add -A is not safe. Stage by path.

Order, revised

Ascending blast radius, unchanged in spirit:

1. Schedule more than one plate, and price a plate-moving arm honestly #3 smallest, additive
2. Name the number every other layer is instrumental to #2 additive, independent of everything else
3. Compose vision with gates, and cover what a workcell needs to keep running #5 additive
4. Printed fixtures, and the path from video capture to a trusted arm policy #6 retarget its base to main once Compose vision with gates, and cover what a workcell needs to keep running #5 lands: gh pr edit 6 --base main
5. Check the ambition against the published record, not just against this lab #4 additive

All five touch README.md, __init__.py and cli.py. Every conflict there is additive: keep both sides. __all__ is sorted, so a merge that drops an entry shows up as:

python -c "import autonomous_lab as a; print([n for n in a.__all__ if not hasattr(a,n)])"

Still open from the original post

The duplication inventory is unchanged: codex/evidence-learning-kernel and laboratory-intelligence-layer remain unmerged with their own samples.py, gates.py, criteria.py, permission.py and record.py. Nothing in these five PRs adds a new duplicate, because each composes the existing layers rather than reimplementing them.