Record the Phase 3 (linear arm) descope and retract the S10 erratum

[jaredgalloway]

What

EPIC #290 will no longer deliver its Phase 3 linear (Identity) baseline arm.
The work has been migrated to #293 as a standalone issue. This PR records
the consequences in HANDOFF.md, the document a fresh maintainer reads cold.

The issue-surface changes were made directly on GitHub and are already live:

Issue	Change
#293	Retitled to a standalone name; body retains the full spec verbatim with a provenance banner (§0) and five epic-coupling statements rewritten. Detached as a sub-issue of #290; no longer blocks #296.
#290	Scope-change banner; Phase 3 section marked DESCOPED; DAG redrawn with the 1 → 3 edge severed and 3,4,5 → 6 reduced to 4,5 → 6; S10 moved to the carried-over-unchanged list; action item 4 marked not-delivered; the S10 erratum landmine corrected in place.
#296	Unblocked (all three surviving blockers closed). Banner instructing that a missing S10 PDF is now correct, not an upstream bug. S10 moved to the carried-over list; the linear-vs-sigmoid gap removed from numbers-in-play.
#297	The instruction to surface the S10 erratum "regardless of what earlier phases decided" is withdrawn, with the supportable replacement sentence spelled out.

Why the erratum retraction is in scope

HANDOFF.md, #290, #296 and #297 all stated that legacy notebook cell 103
plots the sigmoid collection's CV loss inside the linear-model figure, making
published Figure S10 an erratum against the preprint.

The claim is false. #293's spec work recovered the notebook at
fc89753:notebooks/spike-analysis.ipynb — the previously cited 6c98b7b does
not resolve in this repo — and read it by 0-based cell index:

idx	exec	what it actually does
103	127	the linear fit call, not a loss call
105	132	cross_validation_df = linear_mc.get_conditional_loss_df() — linear, and read
106	133	renders shrinkage_analysis_linear_models

Execution counts are monotone 128 → 132 → 133, so the last write before S10
rendered was the linear one.

This mattered enough to fix now because the descope removes the safety net:
with no linear reproduction run left in the epic, nothing downstream would
have caught the claim before it reached the manuscript. Telling a co-author a
published preprint contains an erratum it does not contain is a worse failure
than reporting nothing.

The evidential limit is stated honestly in both #293 and HANDOFF.md: stored
execution counts record an execution order, not proof the saved PDF came
from it. But "unverified" is not "erratum", and only the weaker claim is
supportable today.

Verification

• git worktree-isolated branch; HANDOFF.md is the only file touched (+69/-12).
• Phase 4's landed status confirmed against merged PR Add the naive per-condition baseline arm and manuscript Figure 3 #311 and the presence of
  naive_baseline.ipynb, test_naive_baseline.py, spike_naive_baseline.nblink.
• Confirmed no linear-arm artifact exists in the repo — no
  linear_baseline.ipynb, no rule linear_baseline, no spike.linear config
  block, no S10 in FIGURE_NAMES — so the descope required no code removal.
• Dependency graph re-read from the API after editing: EPIC: Complete the spike manuscript analyses — convergence, baselines, and figures #290's sub-issues are
  291/292/294/295/296/297, and Phase 6 of EPIC #290: Figure manifest for the handoff #296's blocked_by is 291/294/295, all closed.

No code, config, or pipeline changes. Docs-only.

🤖 Generated with Claude Code