ums: bank five process learnings from the 2026-08-24 session

[d-morrison]

UMS pass banking five transferable process learnings from a long 2026-08-24 session.

Seven candidates were considered; three were verified as already covered and are not duplicated (see the skip table below), per grep-is-not-coverage.

Entries written

Learning	File	Shape
A SHA pin on a reusable workflow freezes the caller, not the refs it resolves at run time	memories/github-actions.md	New section, placed to correct the bullet immediately above it, which named this exact pin and is now insufficient
A misleading test label also licenses a deletion --- the direction with no mutation available	shared/coding/fact-check-code-logic.md	Extends the existing Misleading label entry with its inverse
The weakest population is the files already open in front of you; derive from history	shared/workflow/derive-dont-enumerate.md	The shape below "a derivation is still an enumeration of one pattern" --- no query ran at all
An alternative no fixture can isolate may be dead, not merely untested	shared/workflow/fixtures-are-not-evidence.md	Inverts the masking-fixture sections: the finding is about the implementation, not the row
An ellipsis is a sentence boundary to the gate	shared/writing/semantic-line-breaks.md	A boundary neither tool is named for, distinct from the lowercase-follower case above it

Verification

Each claim was derived rather than recalled:

• The pin. ai-config#2165 is the fix; the delegation chain and the 340-vs-637-line split are recorded in this repo's own .github/workflows/validate.yml comment. The executed script was fetched at the pinned SHA and confirmed at 637 lines.
• The ellipsis. The gate's own _SENT_BREAK_RE was run over four inputs at the pinned SHA: ... " And, [...] Everything, and one. Two each split; ... continues did not.
• The availability-sampling gap. files in front of, already open in, availability sampl, returns confirmations returned 0 hits across shared/ and memories/.

Checks

• Real CI gate run locally against origin/main at the pinned SHA: No lines missing semantic breaks.
• scripts/check-links.py: 2306 links across 553 files, no breakage.
• scripts/check-context-closure.py: 8 fragments examined, 0 over the per-file cap. None of the five files is an @-import, so the always-loaded closure is unchanged.

Deliberately not recorded

The disclosure rule and the claim wording themselves --- those ship in #2131.

Skipped as already covered

Candidate	Covered by
Local reformatter and CI check pull in opposite directions	shared/writing/semantic-line-breaks.md --- already carries four dated recurrences, including 2026-08-24
tool-mappings.md is generated from tool-mappings.yml	AGENTS.md, "Canonical sources vs generated output" table
A measurement published beside its own deriving command counts itself	shared/writing/fact-check-prose.md, with the same exclude-the-stating-file remedy

[github-actions]

Warning

Claude review skipped — API credential or quota unavailable. No CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY secret is configured, or account API quota is exhausted. Re-trigger the review by pushing a new commit or re-running the workflow once configured/reset. View run

[dem-extra1]

🤖 Fallback Self-Review (Adversarial)

Head commit reviewed: \33173d56\ (branch \ums/session-2026-08-24-disclosure)

Summary & Verification

1. Scope: Documents five process learnings across:
   • \memories/github-actions.md\
   • \shared/coding/fact-check-code-logic.md\
   • \shared/workflow/derive-dont-enumerate.md\
   • \shared/workflow/fixtures-are-not-evidence.md\
   • \shared/writing/semantic-line-breaks.md\
2. Link & Style Integrity: All relative markdown links verified (2,306 links checked, 0 broken). Full compliance with SemBr discipline.
3. Validation Suite: \�alidate-skills.py, \ est_slb.py, \ est_validate_skills.py, and \ est_check_pr_fully_clean.py\ all passed with 0 failures.

Verdict

Verdict: Clean / Ready for merge.