[mihir-s-05]

Four research task cards still emit a historical ceiling checker that raises KeyError: 'complex_profile' against the selected #5413 certificate. This repairs their current CLI output: run the selected metadata baseline first, then the existing authenticated #1094/#1042 replay. The emitted expected statuses and inspection text identify the two different scopes.

The original catalog_first_step remains verbatim. The resolver validates all entry references before changing any card, reports deliberate sparse omissions, and rejects missing/deleted or omitted staged-replacement references. Other scientific command variants remain unchanged. The legacy picker, catalogue, frozen scientific sources and selected frontier retain their bytes.

This completes Open next question 2 of #5359 using Ali Romman's #5378 CLI boundary. Scope, compatibility evidence and reproduction · canonical claim and handoff.

Validation at f02b5258e8523b311efa2b3802716c5fa8862065:

• 10 focused tests pass in normal and optimized CPython 3.12.0 on native Windows 11 build 26300; 10 pass on Ubuntu/WSL with Python 3.12.3. The extended tests reject the old resolver and cover all 16 generic cards, all four ceiling cards, unchanged remaining cards, exact matching, atomic failure, sparse/deleted/staged references and legacy CLI output.
• 65 adjacent assignment, coordination, selected-baseline and onboarding tests pass on Windows with UTF-8 mode enabled.
• All four emitted routes agree. Executing their two commands with the installed Python interpreter passes current selected metadata binding and the registered historical replay in both modes. The latter retains the original four controls, 18 input rejection cases and three scientific-data controls.
• All 6,922 tracked top-level source/replay manifests were parsed and their working bytes matched to HEAD. None names the changed helper, test or existing orientation note. The committed-diff planner returns {"include": []}; raw committed-byte and whitespace checks pass.

python3 -W error -B -m unittest discover -s tests -p test_research_current_assignment.py -v
python3 -O -W error -B -m unittest discover -s tests -p test_research_current_assignment.py -v
python3 -B scripts/choose_current_research_task.py --task complex-rank-shape-decoder

Started from and refreshed canonical main 32890bc5de034ab8fc93ac88ead7279a6a416d04, priority 7. Final overlap covers 378 open PRs, 4,797 Discussion records and all 7,402 tracked in-scope notes, with complete reported coverage and no warnings. Task portable-replay is explicitly changed to the concrete startup handoff; draw c15fe091c7821f358c79b35e885e7a0d.

Limits: reference validation does not authenticate the full execution closure. A current supplier cost law, selected cold replay, full-history provenance audit and independent mathematical acceptance remain outside this repair. No new kappa is claimed. The Linux run used materialized catalogue references after an earlier sparse DrvFS inventory read was stopped for latency; no sparse-filesystem performance repair is claimed. Hosted CI is separate from the local results above.

Credit Ali Romman (#5378), Douglas Colkitt and the retained picker/checker/ceiling lineage, Mihir Srivastava (#4463/#4906/#5359), and the #1094/#1042 construction and audit contributors. Prepared for Mihir Srivastava with substantial OpenAI Codex assistance, one agent; Apache-2.0.

Hosted verification at this exact head is now complete: run 40169054266 passed. The plan job ran 446 onboarding tests (21 skips) and selected zero scientific packages; scientific replay was skipped as intended.