KJ5HST · KJ5HST · Sep 16, 2026 · Sep 16, 2026 · Sep 16, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -35,6 +35,35 @@ Reverse-chronological, newest on top; prepend-only. Promote to `## YYYY-MM` sect
 
 ---
 
+### 2026-09-16 · [ad hoc] Parallel-sessions plan — making fan-out to concurrent writing sessions clean, not just safe
+
+- **Action:** planning session S24 on branch `docs/parallel-sessions-plan`. The operator asked whether the
+  quality-ratchet work (PR #82) made multi-agent fan-out workable; the Phase 0 answer was *safe but not
+  clean* — `ITERATIVE_METHODOLOGY.md` §Mechanical Gates binds every actor's output, but two writing
+  sessions still collide on `CHANGELOG.md`/`HANDOFFS.md` (prepend-only, one anchor, co-staging hook forces
+  every commit onto them — S21 hit it with two sequences), on serial `S<N>` identity, on the absent
+  merge-session receipt (the open "11 reconciled receipts" ruling), and on Test 9's `--source=github`
+  coupling to `main`. Deliverable: `docs/planning/parallel-sessions-plan.md` + a PR for review, nothing
+  implemented (S13's shape). Results appended at close-out.
+- **The plan, committed:** `docs/planning/parallel-sessions-plan.md` — 468 lines; §1 decomposes the operator's
+  own records into six mechanisms (two-writers-in-one-tree, the return-content fan-out that worked, the S21
+  double-ledger conflict, the undischargeable "11 receipts", Test 9's `main` coupling, the worktree-blind
+  calibrate); §3 the finding (*one closer per tree*; two shapes by who closes out); §4 fifteen decisions D1–D15
+  plus an alternatives table (changelog fragments considered and deferred); §6 six phases, one per session;
+  §7 six honest ceilings; §8 twelve operator decisions; §9 the evidence commands, all re-run before commit.
+- **PR opened (non-commit action):** [PR #83](https://github.com/KJ5HST/methodology/pull/83) from
+  `docs/parallel-sessions-plan` at `903d724`, read back from the API (OPEN, +492/−0, 3 files); its body carries
+  the summary and asks the operator to answer §8 there — that answer is the plan's Phase 0. **Not merged.**
+- **Gate run at `903d724`:** `quality_ratchet: 8/10 pass · 2 fail · 0 unmeasured · results bcbd7f39383a ·
+  manifest 97a7aab85b9a`. The two fails are one **environmental** failure — `tests-sh-passed` 138 /
+  `tests-sh-failed` 1 from `tools/test_context_budget.py` `TestFitGateEndToEnd::test_an_admitting_floor_prints_the_constant`:
+  this machine now holds exactly 4 transcripts for this path (calibrate()'s fit minimum; this session's is the
+  4th), so the fit runs on 4 points and is refused for a negative slope (R² 0.049), a refusal the test's `setUp`
+  does not skip on (it skips only on "not enough" — S16's fix for the sibling case S19 reported). Tool and test
+  are byte-identical to `main @ 6b29d3d`; not loosened; the one-file fix is the next session's first small task.
+- Session S24: claim `3c244aa` + plan `903d724` + the close-out commit (receipt complete, cites the run). Every
+  commit ran the ledger co-staging hook clean — no `--no-verify`.
+
 ### 2026-09-16 · [ad hoc] PR #82 merged — post-merge verification on main and the first tightening
 
 - **Action:** the operator merged [PR #82](https://github.com/KJ5HST/methodology/pull/82) (quality ratchet,

diff --git a/HANDOFFS.md b/HANDOFFS.md
@@ -7,6 +7,22 @@ This repository dogfoods its own methodology: every session records a durable, m
 
 ---
 
+```handoff
+session: S24
+date: 2026-09-16
+status: complete
+self_score: 8
+predecessor_score: 8
+active_task: Planning session — write docs/planning/parallel-sessions-plan.md: how the methodology supports fan-out to multiple concurrent WRITING sessions (the two prepend-only ledgers + the co-staging hook conflict at one anchor; serial S<N> identity; no receipt for a merge; Test 9 couples branches to GitHub main). Evidence-based inventory, per-phase DONE/verify/STOP, open decisions for the operator. ONE DELIVERABLE (the plan document + a PR for review, S13's shape). Nothing implemented. Operator trigger: "I want that plan." COMPLETE — plan committed, PR #83 open; NOT merged.
+what_was_done: Orient (read-only 2.11.1 scan via collect_all, not the 2.6.1 oversight copy; both ledger frontiers = HEAD; S23's gate citation matched --status exactly) → answered the operator's fan-out question from the corpus with file:line (safe, not clean) → routed "I want that plan" as a planning session, asked for /effort max, claim 3c244aa on docs/parallel-sessions-plan → inventory: ARCHITECTURE_WORKSTREAM + the quality-ratchet plan as shape precedent; greps over the runner, both ledger seeds, check-handoff, the hooks, the ratchet, context_budget, the dashboard's ledger-lag, bin/sync + Test 9, HOW_TO_USE, the campaigns; the operator's real fan-out records found in an adopter project (S163: six mutating verify-agents in one tree; S13: return-content agents + serial integrator) — cited by session number, project unnamed → plan 903d724 (468 lines: §1 six mechanisms, §3 one closer per tree / two shapes, §4 D1–D15 + alternatives table incl. changelog fragments deferred, §6 six phases, §7 six ceilings, §8 twelve decisions, §9 evidence commands) → every §9 command re-run before commit; the verify pass found and fixed 9 shifted citations (find_root 61→72; CHANGELOG/HANDOFFS anchors moved by the claim commit, now cited at 6b29d3d; HOW_TO_USE range; README live count claim vs a historical one; dashboard _fold_history 2040→2074) and 2 overstated "mechanical" sentences → pushed → PR #83 opened and read back (OPEN, head 903d724, +492/−0, 3 files) → this close-out. Learnings: none appended — the candidate row ("a tree has one writer; many hands, one closer") is the plan's D1/D2 and becomes Learning #17 in Phase 2 only if ratified; writing it now would implement a decision the plan puts to the operator (FM #18). Stated rather than left unsaid.
+next_steps: (a0) FIRST, small, one file: tools/test_context_budget.py TestFitGateEndToEnd.setUp (:339–357) skips only on "not enough"; calibrate() has a SECOND data-shape refusal — a negative slope prints "no constant recommended" and returns WARN even at floor 0.0 — so on a machine with exactly 4 transcripts (the fit minimum) test_an_admitting_floor_prints_the_constant (:372) FAILS instead of skipping. Fix: probe at the admitting floor too and skipTest when the tool refuses for slope sign; RED first on this machine, then green. It is the sibling of S16's "not enough" fix (S19 gotcha 1). Until then bin/tests.sh reads 138/1 and `--run` 8/10 HERE — environmental (tool + test byte-identical to main; only ~/.claude/projects transcripts differ), and NOT a reason to loosen tests-sh-failed (SAFEGUARDS ratchet row). (a) OPERATOR: answer the twelve §8 decisions on PR #83 (a comment suffices — that is the plan's Phase 0); each carries a recommendation. (b) Per D12, the v3.8 RELEASE PR comes before any phase of this plan (S23's next_steps (a): README What's New in v3.8 — a FRESH entry, the "Since v3.6 (unreleased)" section no longer exists; CLAUDE.md §Versioning paid by reduction against its 23,483-token pin; tag + release). (c) Then Phase 1 of the plan, one session: .gitattributes (four merge=union lines) + starter-kit/gitattributes seed + bin/_manifest.py row (29→30) + a two-branch merge test in bin/tests.sh, RED-first; then bin/check-ledger + an 11th gate — file list and DONE/verify at docs/planning/parallel-sessions-plan.md:300–321. Phase 1 is independent of the release and of the prose decisions; it may land first if the operator prefers. (d) Still open from S23: refresh /Users/terrell/code/methodology_dashboard.py (2.6.1) by ONE-FILE copy; the .context-budget-history.jsonl tracked-or-ignored ruling. The "11 reconciled receipts" and "S19 ledger reorder" items are now DECISIONS 5 and 3 of the plan's §8, not floating rulings. DERIVED: `git merge-tree --write-tree --name-only origin/main origin/docs/parallel-sessions-plan` at this close-out exits 0 — origin/main is 6b29d3d, the branch's base — and stays clean until main moves; any main-side prepend to either ledger before merge recreates the S21 conflict this plan exists to fix (resolve as a union, S24 above S23).
+key_files: docs/planning/parallel-sessions-plan.md:21 (§0 the answer), :42 (§1 six mechanisms), :72 (§3 the finding), :110 (§4 D1–D15), :254 (alternatives table), :289 (§6 phases; Phase 1 at :300), :391 (§7 ceilings), :414 (§8 the twelve decisions), :439 (§9 commands); PR #83 https://github.com/KJ5HST/methodology/pull/83 (body = the summary + the decisions ask); HANDOFFS.md:10 (this receipt); CHANGELOG.md:38 (the S24 entry); starter-kit/SESSION_RUNNER.md:175,177 (the reduction candidates D10 sizes: 199 B, 1,701 B); starter-kit/HANDOFFS.md:82–87 (identity paragraph D5 amends); bin/tests.sh:118 + bin/sync:78 (Test 9 / read_github, D8's target); starter-kit/context_budget.py:867 (D9's target); adopter evidence paths: agent memory `adopter-fanout-evidence` (deliberately not in this public repo)
+gotchas: (1) CITE AT THE BASE SHA: the 1B claim commit prepends to CHANGELOG.md and HANDOFFS.md, so every line number in those two files shifts (+11/+7 here) between Orient and the deliverable — the plan cites them at 6b29d3d and says so; a citation made after the claim at the branch tip is wrong for anyone reading main. (2) `context_budget.py` bare and `--json` runs WRITE .context-budget-history.jsonl (untracked, not ignored) — remove before commit; `--precommit` does not. (3) The plan is NOT distributed (canonical-only, like the ratchet plan) — no bin/_manifest.py row, no sync; check-links does not cover it (its citations were verified by hand and by §9). (4) GitHub's honoring of merge=union is an ASSUMPTION the plan defers to Phase 1 (§7 item 3) — do not state it as fact when answering §8. (5) The adopter project the plan cites is unnamed on purpose; keep it that way in PR comments and in Phase 2's prose (memory: adopter-fanout-evidence). (6) The ratchet is live on this branch: Test 9 stays green here (no manifest row added), `--run` ~3 min. (7) Two-table trap (S12 gotcha 2) unchanged.
+runtime_smoke: n/a as an application launch — the deliverable is a document. Build-equivalent: quality_ratchet: 8/10 pass · 2 fail · 0 unmeasured · results bcbd7f39383a · manifest 97a7aab85b9a — run at 903d724 (the plan commit; ~3 min). The two fails are ONE environmental failure: tests-sh-passed 138 / tests-sh-failed 1, from tools/test_context_budget.py TestFitGateEndToEnd::test_an_admitting_floor_prints_the_constant — this machine holds exactly 4 transcripts for this path (this session's is the 4th), calibrate()'s fit minimum, so the fit RUNS on 4 points (n=4, R² 0.0493, slope −3.57) and is refused for a negative slope, a path the test's setUp does not skip on. Reproduced directly (rc 1, "no constant recommended"). Same tool + test at 6b29d3d (git diff --quiet: identical); the other 8 gates pass at their measured values; check-handoff-all 0; check-links 0; check-learnings 0. Not loosened; fix is next_steps (a0); bin/check-handoff --all --allow-pending OK at claim; the plan's §9 evidence block re-run in full before the plan commit (every output matched the plan's claim; 9 citations corrected as a result). The PR read back through the API (state, head sha, additions), not inferred from the CLI.
+changelog_ref: CHANGELOG.md "2026-09-16 · [ad hoc] Parallel-sessions plan — making fan-out to concurrent writing sessions clean, not just safe" — written at claim, one bullet per checkpoint, completed here with the PR
+commit: 3c244aa (claim) + 903d724 (the plan) + this commit (close-out); PR #83 opened at 903d724 (non-commit action, in the ledger)
+```
+
 ```handoff
 session: S23
 date: 2026-09-16