devdocs/developer/ is 83 unowned pages, 73 of them untouched since the day they were created

• Type: task (documentation audit) — the sweep, not the fixes
• Track: A — CLAUDE.md places the internal dev docs with A/B. Per-page findings route to whichever lane owns the subject; this ticket is the sweep.
• Found: 2026-08-30 by frankD, after auditing all 42 live devdocs/dev/** references and all 41 docs/** pages.

Why this tree and not another — the rationale is a measurement

The night's audit produced one finding that says where to look next, and it contradicts the assumption the fleet had been working under:

Accuracy tracked who is accountable for the page, not how many people read it.

docs/** is owned by Track D, is published to the website, and has fewer readers who could check it against the tree — and it was the more accurate of the two sets. devdocs/dev/** is read by every agent every session and is owned by nobody in particular, and it carried six superseded gate rules, four completed-work-described-as-open claims on one page, and a dispatch model describing worktrees that do not exist.

The many-eyes assumption predicts the opposite. Readership did not correct errors; in the worst case it preserved them, because the stale rules were all tighter than the rules that replaced them, so obeying one cost ten minutes and produced nothing wrong. Nobody reports that.

So the next sweep should be ordered by which pages nobody owns, and devdocs/developer/ is the largest such surface in the repo.

The measurement

CLAUDE.md names docs/** (D), devdocs/dev/** and devdocs/progress/** (A/B). devdocs/developer/** is named nowhere, by either document that assigns documentation ownership. It is unowned by construction, not by neglect.

The citations broke in the commit that created the tree

Extracting every compiler|lib|tools|test|docs/... path mentioned across the 83 pages gives 157 distinct paths, of which 35 do not resolve. Honest breakdown, because the raw number overstates:

• 9 are pure docs/ → devdocs/ renames — the file is right there under the new prefix.
• ~3 are prose placeholders, not citations (compiler/name.pas, compiler/name.c, compiler/name.h in a "name your file" sentence). A path extractor cannot tell those from a real reference; a reader can.
• ~23 are genuinely stale, including compiler/parser.inc (split into pasparser_*.inc on 2026-08-20), docs/todo.md, docs/limitations.md, docs/pascal-dialect.md, lib/rtl/anybox.pas, lib/rtl/sqlitedb.pas, and five docs/progress/** ticket paths whose tickets have since changed directory.

The sharpest one: project-state.md cites docs/historic/direct-codegen-legacy.inc. That file exists — at devdocs/developer/historic/direct-codegen-legacy.inc. The commit that broke the citation is the commit that created this tree (601c08105, "docs: split public and internal docs", 2026-06-26), and in the 65 days since, nobody has followed the link.

A page whose heading asserts currency

project-state.md opens "Project State Audit / Audited: 2026-06-12" — 79 days ago — and its first architecture bullet reads:

"Pascal, C-subset, and BASIC-subset frontends share the native x86-64 Linux ELF emitter."

Since then Nil Python became a mainline frontend with its own gate, Rust and Zig arrived as experimental ones, and the emitter grew to six code-generating targets. A reader taking that page at its word gets a 2026-06 project.

This is the same shape found twice already tonight — Roles — LIVE, ## Current assignments, a ## TODO over five done/ tickets. A heading is an assertion that nothing re-derives, and Audited: <date> is the strongest form of it, because it explicitly claims someone checked.

Name collisions with the pages that ARE owned

devdocs/developer/ contains architecture.md, cli.md, compatibility.md, c-interop.md, concurrency-memory-model.md. docs/reference/ contains architecture.md, cli.md, and status.md covering compatibility.

Two files with one name, one owned and current, one unowned and two months stale, is the glossary.md defect at directory scale: two definitions of one thing is not redundancy, it is a coin flip, and the reader does not know they tossed one.

What this ticket asks for

Not "fix 83 pages". In order:

1. Decide the tree's status — this is the real question and it is a Track U call if A does not want it. Three options, and they are not equal:
   • live reference — then it needs an owner and a README.md classifying its files, the way devdocs/dev/README.md does. That page is the mechanism that made the 42-file sweep possible at all.
   • historical record — then say so at the top of each page and stop, because a record is not wrong for being old, and correcting one falsifies history.
   • split — most likely correct: project-state.md and todo.md want to be one or the other, and the historic/ subdirectory shows the split was already begun.
2. Then sweep whatever remains classified as live. The tooling exists: tools/docaudit.py (cites, slugs, limits, targets, comments) is pointed at devdocs/dev and needs a directory argument, which is a small change and the cheapest part of this ticket.

The sweep itself — measured 2026-08-30, audit only, nothing edited

The ticket above was filed from an inventory. The sweep then ran, and it found something sharper than staleness. Recorded here because it changes what the classification decision is for.

1. The tree hands agents the exact command the repo refuses

Four sections in unowned pages are headed as gates and open with a command that .claude/hooks/no-full-suite.sh now denies:

All four repaired 2026-08-30 by frankD, under a grant extended to exactly these files and no others. Each now carries the per-fix loop and a pointer to CLAUDE.md as the authority, with a dated Superseded note saying what the instruction used to be. threads-todo.md was annotated, not cut: items 2-4 of its ladder are still correct and only item 1 moved, so deleting the item would have destroyed the information that the rest still stands. esp32-support.md keeps the make test-esp-bare sentence as a description of what the target does — only the prescription went. Nothing else in the tree was touched, and the rest still waits on step 1.

The hook's own header comment names the reason it exists: an agent "reached for make test-nilpy twice in one session." These pages are where that instruction still lives. The hook is in .claude/ and is owned; the instruction is in devdocs/developer/ and is not, so nobody connected them.

This is not a stale fact. It is a live supply of the behaviour the hook was built to stop, and an agent following it hits a REFUSED wall with no idea why, because nothing on the page knows the rule changed.

threads-todo.md's ladder is the autonomy.md shape again — partly stale. Items 2 and 3 (self-host fixedpoint; the same path under --threadsafe) are still exactly right. Only item 1 is superseded. A reader cannot tell which.

Honest scoping, because the raw grep overstates. 24 of 58 top-level pages mention a now-denied command; only these four prescribe one. The rest are descriptive — "make test covers procedural Pascal, units and arrays", "all wired into make test", and cli.md's reference table of what each Makefile target does. Those are facts about where coverage lives, they are not wrong, and they should not be swept. project-state.md:155 looks prescriptive in a grep and is not: it sits under "Verification passed after the batch:" — past tense, a record, correctly.

2. Three "not implemented" claims, disproved by measurement

Compiled against $(PXX_STABLE), not reasoned about:

• todo.md:306 — "Float intrinsics. Trunc, Round, Int, Float not implemented", and limitations.md:62 says the same. All work: Trunc(2.7)=2, Round(2.5)=2, Int(2.7)=2.0, Frac(2.75)=0.75.
• nil-python.md:77 — "Containers, classes, dynamic attributes ... are outside the v1 frontend", and nil-python-plan.md:126 — "containers, no classes yet". A nine-line .npy with class Point and a method compiled and printed 7. NilPy is mainline with a gated suite; CLAUDE.md lists classes in its coverage.
• ir-handover.md:104 — "Optimization passes (none planned yet — this is by design for now)". Track O is a formal lane; -O2 is the proven default and -O3 a free tier for new passes.

Same class as concurrency.md's false limitation found in docs/** last night, and the same mechanism: a "not supported yet" is never contradicted by use, because nobody who reads it tries. Optimistic errors get removed by the first reader who hits them; pessimistic ones accumulate forever.

3. The heading that asserts a legal position

licensing-concerns.md:10 — ## Current Position: No License Yet.

The repo root now carries LICENSE and LICENSE.md, and README.md states a per-component split: zlib for the embedded builtin code, 0BSD for examples, CC BY 4.0 for docs. The page carries no **Status:** line, so nothing marks it as a record of a past position.

Flagged, not fixed — but of everything in this tree this is the one whose staleness is not merely an engineering cost, and a launch is the moment someone reads it.

4. Why nothing here was corrected — the tree does not classify itself

12 of 58 top-level pages carry any self-classifying line (**Status:**, **Updated:**, **Audited:**, **Authored**). Seven more declare themselves by filename prefix (plan-*, anomaly*), and the 34 files under historic/ and anomaly-evidence-2026-06-02/ declare themselves by location. That still leaves roughly forty top-level pages a reader cannot classify without reading the git log.

The best of them show it is a solved problem when someone bothers:

• plan-rtti-streaming-lfm.md:3 — "Status (2026-05-31): delivered. This document is retained as the design [record]". Perfect: old, accurate, and says which.
• frontends-and-targets-strategy.md:122 — "## Current read (subject to change — this is brainstorm)". A currency heading that disarms itself.

And the worst is the one the ticket already named: nil-python-plan.md is titled "Plan:" for a frontend that is now mainline with its own gate, and says nothing anywhere about having been delivered. The plan was executed; the plan does not know. A reader cannot tell a plan not yet done from the record of one that was — and that ambiguity is what makes correcting any of these pages unsafe, because correcting a record falsifies history.

So: nothing in this tree was edited. That is not caution, it is the finding. The classification in step 1 below is not preparation for the real work — it is the work, and roughly forty pages need one line each.

Revised recommendation

Step 1 of the ask stands, and gets cheaper and more urgent:

• The four gate prescriptions above were separable and did not wait for the classification — they are done, see the note under that table.
• Scope check for whoever picks this up. The grep that finds those four overstates by 6x: 24 of 58 pages mention a denied command and only four prescribed one. Do not run the sweep on the grep. And the grep itself trips the hook, which scans command text and does not care that the command sits inside a search pattern — a REFUSED there is an artefact of the search, not a finding.
• The licensing-concerns.md heading is a Track U call, not a docs edit.
• Everything else waits on classification, and the cheapest sufficient classification is one **Status:** line per page — not a rewrite.

Prior art from the sweep that produced this

• devdocs/dev/README.md §4 — where that directory rots, and why a command attached to an obligation is necessary but not sufficient.
• The rot classes worth grepping for first, ranked by what they cost: superseded operating rules (all six found were tighter than current policy), finished work described as open, headings asserting currency, and dead ticket/file citations.

Gate

None — audit and classification. Any code or library defect the sweep finds is filed into the owning lane, never fixed under this ticket.