[mihir-s-05]

Change and mathematical scope

Sixteen of the 22 task cards still emit python3 -B scripts/research_baseline.py. On the authenticated selected #4470 inputs, that helper raises KeyError: 'bit_lower'. Current priorities already name the compatible successor, but the picker does not expose the catalogue's selection context.

This documentation-only contribution puts an explicit generic-command substitution in the standing procedure and links schema-aware sparse diagnosis. It also adds a reproducible, source-pinned dependency audit explaining why the historical catalogue, picker and baseline are preserved.

The audit reads all 5,854 top-level SOURCES/VERIFY manifests at canonical base 624de9dc66858af0945e5f8a0942488d62280d43. The catalogue has 16 live input pins, 2,676 provenance rows and 57 declared replay dependents; the procedure has zero live input pins or declared dependents and three provenance rows. All 16 live catalogue pins already differ from current main: this is not a claim that the proposed change newly breaks 16 passing packages. Actual planner results are zero packages for this change, versus 57 for a hypothetical catalogue edit.

Only docs/research/agent-procedure.md and the new docs/research/task-startup-compatibility.md change. No scientific source, receipt, helper, card, allocation weight or selected-frontier file changes. No new κ or mathematical acceptance is claimed.

Task and overlap

• scope-navigation, refined to one broken startup path; scope key b18e0e173a4430a9564b, nonce cc0ba7957f5aa29a9b8e5a07a7a6ec93, priority revision 7.
• Canonical claim and migration handoff. The dependency audit narrowed the implementation from editing pinned cards to fixing the active procedure.
• #4463, head 108509ac7a441617e3783fbd991bdf249fd23e64, supplies the existing productive-helper adapter. #4631, head 48aef2349dc761f6b3b6bf3e59c3ee8080dbbe7a, supplies sparse schema selection. This change connects those tools to the card-following procedure.
• Canonical main remained 624de9dc66858af0945e5f8a0942488d62280d43 at the pre-submission refresh. Complete cached pagination refreshed 425 open PR bodies and 3,957 Discussion records at 23:32–23:33 UTC. No matching active migration surfaced. Sparse local README/TASK coverage, arbitrary diffs and undeclared runtime reads are not covered by that text search.

Validation

At commit 6b179376a3fe18a6efc57951c2df48bef9fed117, native Windows, CPython 3.12.0:

• Actual selected source closure authenticates; ordinary and optimized successor reports are byte-identical (PASS_RESEARCH_BASELINE_BINDINGS_ONLY, κ 60403/20000000). The unchanged legacy command reaches its expected post-authentication bit_lower failure.
• 13 selected-baseline controls, 16 assignment controls and 8 coordination controls pass. The coordination suite needs -X utf8 on this host; its default cp1252 run has one pre-existing decoding error in unchanged autoresearch.md.
• The note's exact-Git-object snippet reproduces the inventory independently of the working-file/Git-blob equality audit. All 5,854 metadata files matched their committed blobs.
• Both changed files match committed bytes; links and git diff --check pass. The standard changed-package planner returns {"include": []}.

python -B scripts/research_selected_baseline.py
python -O -B scripts/research_selected_baseline.py
python -B -m unittest discover -s tests -p test_research_selected_baseline.py -v
python -B -m unittest discover -s tests -p test_research_assignment.py -v
python -X utf8 -B -m unittest discover -s tests -p test_research_coordination.py -v
python -B scripts/research_package.py --changed-since 624de9dc66858af0945e5f8a0942488d62280d43 --matrix-batch-size 256

No broad historical, native/selected cold, or Lean replay was run. Hosted run 39872761723 passed at the exact submitted head: onboarding ran 346 tests in 111.101 seconds with 11 skips; package planning returned {"include": []}, and scientific replay was skipped as intended. This is automated tooling validation, not mathematical acceptance.

Credit and remaining work

Credit Douglas Colkitt and the retained onboarding/source-checker contributors; Mihir Srivastava / #4463; lee-phys / #4631; Mister-Meeseeks and the credited Claude agent / #3725 for the immutable-helper precedent. Scientific attribution remains with the selected contributor map. Prepared by Mihir Srivastava (@mihir-s-05) with substantial OpenAI Codex assistance; no subagents. Apache-2.0.

The emitted historical cards remain unchanged. A direct consumer that skips the procedure still sees the old generic command. The next bounded task is a migration policy for the 57 declared dependents, distinguishing existing drift from new incompatibility and preserving original task/source history. Nested manifests, archives and hidden runtime dependencies remain outside this audit.