docs(phase-33.5): S90 checkpoint — decision-log entry 'P33.5 S90' (the re-charter, the three audits + the adversarial review, the six owner decisions, what tasks 0–9 measured, the hindsight path); CURRENT_PHASE checkpoint refreshed for a fresh session (tasks 0–9 done, NEXT = task 10 at Max; the decision-log item of task 14 already banked)

This commit is contained in:
Drew T
2026-09-07 18:12:31 -06:00
parent caefe98728
commit e1463d4300
2 changed files with 53 additions and 3 deletions
+43
View File
@@ -3504,3 +3504,46 @@ register lever (accelerators (16)), and treat "PROVED" as "proved against this l
- **Hindsight path.** Ask before offering: "what does this maintainer's workflow need?" — an interactive tool's author does not want a
mode for unattended search; a short issue describing the floor with the two numbers, in Drew's voice, asking whether a symbol
rename would be the intended fix, would have got the same technical answer at a tenth of everyone's cost and no ill will.
## P33.5 S90 (2026-09-07) — the docs sub-phase before the flip: the wiki becomes the source of truth, the tracked scratch is pruned, and the day-one kit is chartered as a Phase-0.5 overlay on ProjectArchitect 2.0
- **Context and belief.** Phase 33 closed at v1.32.0 with the flip gated on GitHub Support; the wiki (13 pages + 13 chapters) had
been written in P33 F3 as a summary layer OVER `docs/`, linking back into it. The belief at the P33 close was that the docs tree
was publishable as it stood. Drew's review of the folder on 2026-09-07 found otherwise: ~100 authored files from 33 phases —
a hindsight study written at 78% with a commission nobody had fulfilled, two distillation ledgers whose banners said "never
applied" (both HAD landed, as cookbook §265–§269), a 216 KB raw agent-research JSON, eleven per-session tool designs, frontier
reports and worklists for sessions long closed — beside the live references; no page on how `docs/` or `.run/` are used, no
gitignore template, no AI-conduct rules on the wiki, nothing on Gen3; 1,086 tracked `.run/` files visible on GitHub, 172 of them
tracked by index inertia alone; and 81 memory files under `~/.claude` that no future project could inherit.
- **The three audits (read-only agents) measured the gap** rather than guessing it: every `docs/` file classified (45 KEEP-LIVE, 12
generated, 6 INTEGRATE, ~60 SUNSET incl. 6 frozen frontier snapshots and 12 orphaned per-binary reports); the `.run/` tree by group
with its live readers (and one firewall gap — two tracked disassembly listings of one matched function, invisible to a path-and-hash
audit); the memory store by class (52 already in the repo, 14 portable-not-in-repo, 9 project-specific, 7 stale, 1 off-project).
An adversarial review of the plan then caught six ordering and design hazards before execution — the forward-link race with the
link checker, the purge set doubling as the history gate's census, regenerating snapshots into empty documents, a content-check regex
that would not have matched its own offenders, a kit that could not write into a constitution ProjectArchitect had already frozen,
and a dry-run that could write into the real tree.
- **The pivot (Drew, 2026-09-07).** Re-charter the next work as **Phase 33.5**, a docs sub-phase BEFORE the flip: (1) the wiki is the
single source of truth — every `docs/` file is KEEP (listed in a Reference index, the only sanctioned wiki→docs link site),
INTEGRATE (folded into a page) or SUNSET (moved with history into `docs/sunset/`, indexed, for Drew's review — deletion is his);
in-repo documents link wiki pages, not `docs/` files; (2) prune only the TRACKED `.run/` (what GitHub shows) — the untracked
29 GB is out of scope; (3) the wiki carries the conventions (docs, scratch, the ROM firewall with a copyable gitignore); (4) the
day-one kit `decomp-architect/` is built IN-TREE now and split out after the flip as xsig was — and, from the review, it runs AFTER
ProjectArchitect 2.0 as the project's Phase 0.5, shipping an intake document for PA's own interview instead of writing the
constitution; (5) the memory store is reconciled but stays outside git (kit seed only — no `.claude-state/` retrofit for a repo
about to go public); (6) the two disassembly listings are untracked and named in an audit-only rules file, with NO second history
rewrite — the tracked C and the pinned compiler reproduce those bytes, and another force-push would add another set of old tips to
the open Support ticket. Rules R74–R83 (the P33 candidates) ratified at the gate.
- **Measured through task 9 (S90):** ten commits; 17 wiki pages + 13 chapters, all reachable; the six INTEGRATE documents folded
(the hand-matching guide's five signature moves, the wave ledger's per-wave prices and instruction-weight metric, the portable
workflow's six residual items, the hindsight study's permuter-failure track); 59 files archived with history; `doc_links` grew
from one check to six (archive refusal, the index-derived allow-list, docs coverage 64 of 64, git-based citation classification with a
"(not kept)" declaration, wiki-first warnings); the render selftest asserts reachability; the timeline was found STALE at HEAD and
wired into report/audit-digest; tracked `.run/` 1,086 → 868 with `audit_public` gaining a content check whose first cut the
negative control caught (`nop` broke the run at 60 < 64); the history gate still PASS; the project memory store 80 rows == 80
files with seven stale memories corrected in place; the kit's 16-seed memory pack written with zero project literals.
- **Hindsight path.** Write the wiki as the source of truth FIRST and the summary pages second — P33 built the pages over the
records and left the records as the reference, so the consolidation had to be its own phase. Two instruments that should have
existed from the first wiki commit: a link policy the checker enforces (wiki-first, no links into an archive) and a content check
in the ROM audit — both were one-day builds. And run the referrer census and the pending-list mechanism as a habit for every
document move; the one-shot audits found nothing the checks would not have found continuously.
+10 -3
View File
@@ -452,9 +452,16 @@ WAIT for Drew. Rebuild the harness task list (16 rows, R28) marking tasks 0–9
**Phase 33.5** (sub-phase; v1.32.0 → v1.32.1), gate 1 approved 2026-09-07 by Drew in plan mode at Max; effort follows the plan's
column (Max for tasks 10, 12, 15 — prompt at each transition, R27). Baseline HEAD `80d45b29b`; task 0 = `39d524991`; task 1 =
`a0cf302e5`; task 2 = `d06923a06`; task 3 = `5d10a0d12`; task 4 = `9970f1e62`; task 5 = `a0d4ae836`; task 6 = `6ec4786bd`; task 7 =
`21c98ed5a`; task 8 = `ae71efe56`; task 9 = the commit after it (the seed). No build input changed; the fleet is 218/218 at the
Phase-33 close. Tracked `.run/` = 868; `audit_public` OK; `gate_scan` PASS; `doc_links --strict` green; pending list empty. The kit
directory so far holds ONLY `decomp-architect/memory-seed/` (16 seeds + MEMORY.md). Tasks 10–13 build the rest of the kit; 14–15 close.
`21c98ed5a`; task 8 = `ae71efe56`; task 9 = `caefe9872`; the S90 checkpoint commit follows it (this file + the decision-log entry).
No build input changed; the fleet is 218/218 at the Phase-33 close. Tracked `.run/` = 868; `audit_public` OK; `gate_scan` PASS;
`doc_links --strict` green; pending list empty; tree clean. The kit directory so far holds ONLY `decomp-architect/memory-seed/`
(16 seeds + MEMORY.md). Tasks 10–13 build the rest of the kit; 14–15 close. **Already banked for task 14 (R31, done at the S90
checkpoint):** the `docs/decision-log.md` entry "P33.5 S90" — task 14 must NOT write a second one; it still owes the wiki page, the
README/Tools rows, the SETUP rows for the kit and any accelerators entry. **Session S90 (2026-09-07) ended here at 85% context**;
the next session is S91. Drew's effort at the end of S90 was xHigh (toggled to Max for task 3 and back); task 10 is Max — prompt.
Nothing is running in the background; the SessionEnd hook stops the headless Ghidra MCP (no RE work happened; nothing under
`ghidra/` is tracked). The plan file `~/.claude/plans/max-effort-set-plan-encapsulated-muffin.md` is a copy of the "Approved plan"
section at the end of this file — this file is the one that counts.
### 2. What NEXT does (task 10, **Max**) — exact steps (plan D5)
Read first: `/mnt/z/Storage/git/ProjectArchitect/README.md` (the install ritual, "What gets installed"), `…/project-architect-2.0/SETUP.md`