The free section index is documented, and it is not being used

Track U. Fork of intent in CLAUDE.md's own text; owner's file, so this is filed rather than edited.

The observation

CLAUDE.md says, of devdocs/dev/debugging-playbook.md:

has the tool for your case — LOOK UP THE SECTION. 279KB, ~70k tokens, 72 sections; grep '^## ' lists them free.

Both halves are correct. The first is vivid and carries a number that hurts. The second is a subordinate clause after a semicolon. A rule that stops you paying 70k tokens should not also stop you paying zero, and in practice it does.

Measured, one session, 2026-09-05

Three instances, and the two most experienced readers of this file produced two of them:

1. frank-optimize told the coordinator "no load-independent instrument exists on this box", then searched for valgrind and qemu TCG plugins. The answer was ## \perf` being blocked is not "no profiler" — build the compiler with FPC and `-pg``, playbook line 4798, and in the section index at line 174. It measured the thing that closed [[feature-inline-nonleaf-and-branch-locals]] in eleven seconds with no root.
2. frank-coordinator wrote the same claim into the roster without grepping the file named for the question.
3. frank-optimize reported objdump -d as unable to disassemble a pxx ELF. The correct mechanism (-g, or no section headers are emitted) was verbatim in that session's own notes, in a session where those notes were loaded.

In every case the search that replaced reading cost more than reading would have. Nobody declined to consult the index; nobody considered it, which is the tell that the instruction is being read as a prohibition rather than a price.

The fork

• (a) Reword. Lead with the free index and demote the cost: "grep '^## ' lists its 72 sections free — do that first. Read a section only when it is yours; the whole file is ~70k tokens." Same two facts, order reversed.
• (b) Leave it. The rule is literally correct and the failures are reader error. Costs nothing to keep, and three instances in one night may be a bad night rather than a pattern.
• (c) Something mechanical — a tool that greps the index for you, so consulting it is not a decision anyone has to remember to make.

Recommendation: (a). It is a sentence, it changes no behaviour anyone relies on, and (c) is a larger job that (a) does not block. The argument for (a) over (b) is that the failing readers were the sessions best placed to know better, which makes reader-error an expensive theory.

Explicitly NOT proposed

Do not read this as an argument to shrink, split or restructure the playbook. The file is fine and its size is what makes the index valuable. A restructuring proposal would rank as a project and never happen, and it would not fix the thing measured here — the index already exists and already works.

2026-09-06 — THE QUOTED NUMBERS ARE STALE, AND THE DRIFT ARGUES THE TICKET'S OWN CASE

Every figure this ticket quotes from CLAUDE.md was corrected at fe0c7e2cd:

The playbook TRIPLED while the pointer to it stayed still, and nothing errored — the pointer kept answering, with a number that had been true. That is this repo's own "the name is not the thing", applied to the very sentence this ticket is about.

It strengthens the fork rather than changing it. The ticket's case is that a vivid cost warning beside a subordinate free-index clause makes sessions stop consulting the file at all. At 279KB that was an over-application; at 905KB the cost warning is more vivid and the index is more valuable, because 237 section headers are the only affordable way in. The correction at fe0c7e2cd also downgraded "lists them free" to "lists the sections for a few thousand tokens, which is cheap, not free" — accurate, and it makes the subordinate clause weaker, which is the direction this ticket argues against.

Still a wording fork and still the owner's. What is settled is only that the numbers no longer support an argument from either side, and that any future phrasing should carry a date or no number at all: a size in a pointer is a lower bound with a date on it.