Cite resources at section grain: the pre-principle-32 corpus tail

[m2ux]

Design principle 32. Cite Resources at Section Grain and AP-134 whole-resource-for-one-section landed in #351. This issue registers the corpus that predates them.

The measurement

A citation is a delivery instruction: the id it resolves to — path after resources/, .md stripped, #anchor kept — is the unit the server loads and bundles. foo.md#bar delivers one section; foo.md delivers the file. Resources over the DEFAULT_MAX_EAGER_RESOURCE_CHARS cap (80,000) never eager-bundle, so anti-patterns (137k) costs nothing and is excluded throughout.

119 (technique, resource) pairs cite a resource whole, totalling 590,718 eager characters. #351 closed twelve of the fourteen that were also anchoring sections of the same resource, the class where the waste was literal duplication and provable by arithmetic. What remains needs each technique read to know which sections it consults.

Scope of this issue: the tail

The top twenty pairs by delivered size are being handled on a separate branch. This issue covers the remainder — roughly a hundred pairs, each under ~8,000 characters, where the saving per site is between a few hundred and a few thousand characters.

Ranked head of the tail, for orientation:

chars	technique → resource
8,816	strategic-review/review-scope → work-package/architecture-review
8,807	area-derivation/derive-areas, evidence-probes/probe-area → midnight-system-review/probe-catalog
8,001	area-derivation/amend-plan, area-derivation/derive-areas → midnight-system-review/subsystem-map
6,595	portfolio-analysis → prism/strategist
6,332	write-cicd-report/attach-remediation, write-cicd-report/write-report → cicd-pipeline-security-audit/remediation-playbook
5,766	requirements-elicitation/ask-question, requirements-elicitation/elicit → work-package/requirements-elicitation

Reproduce the full ranking by extracting every markdown link into resources/ from each technique, resolving the delivery id the way extractResourceIds does, and summing the resource body against the sections the citing prose actually names.

What "fixed" means per site

Per AP-134, three outcomes are all valid, and which applies is a judgement about the technique:

1. Cite the sections — the prose names one or a few sections, and they total less than the file. Point each citation at its section, link text the section title.
2. Leave the whole-resource citation — the technique reads most of the body (a filler working a ## Template with the ## Rules that populate it, an audit walking every entry). Principle 32 explicitly endorses this.
3. Split the resource — no single section covers the need and the resource is large. That is principle 30's case, and it changes the resource rather than the citation.

Outcome 2 is common in this tail: many of these resources are small creation guides whose two sections are the whole file. Recording that verdict is as much a result as a rewrite.

Two sites deliberately excluded

review-summary → review-mode (43,781ch delivered) and validate-specification → validation-rubric (3,908ch) both need sections totalling more than their file. The whole-resource citation is the economical delivery there, and the residue is the bundler shipping a section body when its parent file is already in the bundle — server work, tracked with the delivery-ledger change in #356, not a definitions fix.

Guard opportunity

AP-134's second tell is mechanical and needs no judgement: a bare citation in a technique that also anchors the same resource. That is what made the fourteen findable by arithmetic, and it is checkable in scripts/. Worth adding so this class cannot silently return — the first tell (prose naming a single section) stays an audit judgement.