Plan: reconcile the write/read-side spec change with its consumers

Fixes for the five findings from reviewing the uncommitted agent-memory-system-review.md / .schema.yaml change against its decision log (review-framework-design.md D3/D4), the parser, the schema, and the 190 existing reviews. Comment inline; I'll act on your marks.

The gap in one line

The spec was edited to its decided end-state (write side / read side, no read-back "timing"), but its consumers were not moved with it: the parser still has the timing axis, the schema description string still names the old section, and D3 claims work that isn't done. The spec is ahead of the code.

Findings → fixes

Fix 1 — Remove the read-back timing axis from the parser (Finding 1, the real defect)

D3 says the rb_pre_action/rb_post_action columns "are removed" — they are not. The spec now forbids authoring the token, so the axis is dead. Remove it.

src/commonplace/lib/systems_matrix.py:

  • delete the "Read-back timing" entry from ONEHOT_AXES (lines ~55–56)

  • drop "Read-back timing" from PUSH_AXES (line ~70)

  • drop *ONEHOT_AXES["Read-back timing"] from COLUMNS (line ~91)

test/commonplace/lib/test_systems_matrix.py:

  • remove the rb_pre_action/rb_post_action assertions

test/commonplace/lib/fixtures/zikkaron_review.md:

  • strip the **Read-back timing:** token so the fixture matches the new spec

Then rebuild and confirm the two columns disappear:

  • python3 scripts/build_systems_matrix.py (or the documented command)

  • pytest test/commonplace/lib/test_systems_matrix.py

  • verify kb/agent-memory-systems/systems.csv header no longer has rb_pre_action / rb_post_action

Decision point: removing the axis is the decided end-state and makes D3 true. The alternative is to defer — keep the parser as-is and reword D3's consequence to "pending" (matching D4). I recommend removing now; it's bounded and the columns are already meaningless under the new spec.

Fix 2 — Update the schema's stale heading description (Finding 2)

agent-memory-system-review.schema.yaml, headings.description still reads "## Trace-derived learning placement and ## Read-back placement are optional…". Reword to name ## Write-side placement (with ### Trace-derived learning as its sub-section) and ## Read-back placement. Description-only, not enforced, but it's drift inside the edited file.

Fix 3 — Tighten the Operations wording in the spec (Finding 4)

In ## Write-side placement, the Operations bullet says "Note for each whether it is manual or automatic," but the **Curation operations:** token rule is automatic-only. Reword so curation operations are the automatic ones and manual maintenance is recorded as agency only — removing the invitation to tag manual ops that the token then discards.

Fix 4 — Make D3 accurate (depends on Fix 1)

If Fix 1 lands, D3's "are removed" becomes true — no edit needed beyond a last-checked bump if desired. If we defer Fix 1 instead, change D3's consequence to "pending" so the decision log stops claiming finished work.

Out of scope here (tracked, not blocking)

  • 84 active reviews still carry **Read-back timing:** tokens and ~190 carry the old ## Trace-derived learning placement heading. After Fix 1 the tokens are inert prose and the schema relaxation keeps the headings valid. Stripping the dead tokens and renaming headings is retrofit work — fold it into the existing comparative-review-data-refresh / review-template-retrofit passes, not this one. See kb/work/review-template-retrofit/runbook.md for the current layout-retrofit procedure.

  • Wiring **Write agency:** / **Curation operations:** into the parser as one-hot columns (Finding 3 / D4). Authored-in-prose now, extract-later — same precedent as read-back signal. New axes + columns + retrofit; a separate unit.

Suggested commit shape

  1. Commit A — spec + schema (the currently-uncommitted change plus Fixes 2 & 3): the spec and its own schema move together. Self-consistent on its own.

  2. Commit B — parser + tests + fixture + rebuilt CSV (Fix 1): the consumer catches up; D3 becomes accurate.

Splitting keeps the prose contract and the code contract reviewable separately. They can also go as one commit if you prefer a single atomic "spec end-state" landing.

Verification checklist

  • [ ]

pytest test/commonplace/lib/test_systems_matrix.py green

  • [ ]

pytest full suite green

  • [ ]

systems.csv rebuilds with no rb_*_action columns and otherwise no diff

  • [ ]

commonplace-validate clean on a trace-derived review and a push-activation review (schema relaxation still admits old headings)

  • [ ]

no remaining Read-back timing reference in src/ or test/

Open questions for you

  • Fix 1 now (remove axis) or defer (reword D3 to pending)?

  • One atomic commit or the A/B split above?

  • Do the 84-review token cleanup here as a mechanical sweep, or leave it entirely to the data-refresh pass?