Resolve #19: formalize cross-reference completeness as Learning #7 + Phase 3F hook

CLAUDE.md

@@ -84,7 +84,7 @@ Each phase is hard-gated — you cannot skip ahead. The most critical gate is be
 
 ## Versioning
 
-Changes are tracked via git commits and the README's "What's New" section. Current version: v2.7. Key additions by version:
+Changes are tracked via git commits and the README's "What's New" section. Current version: v2.7.2. Key additions by version:
 
 - **v1.0:** Initial 9 principles, 6 phases, 4 workstreams
 - **v1.1:** Protocol erosion detection, ghost session prevention, minimum handoff requirements, 4 new failure modes (#14-17)
@@ -99,3 +99,4 @@ Changes are tracked via git commits and the README's "What's New" section. Curre
 - **v2.6.1:** Docs-only dot release bundling four external contributions (rmsharp PRs #18/#23/#25/#27). (1) **Skill-citation discipline completed** (#18): `/init` actually cited in `BOOTSTRAP.md` Step 4 with a skip path for existing CLAUDE.md projects; `/fewer-permission-prompts` dropped from the index (no natural citation site); new "A skill is not a phase" principle paragraph in `ITERATIVE_METHODOLOGY.md` §Recommended Skills (a skill that pulls a session across a hard gate is FM #2 wearing a tool costume — gates bind regardless of loaded tools); new "Skills not recommended (and why)" 10-row subsection in `RECOMMENDED_SKILLS.md` surfacing the v2.6 audit's deliberate omissions plus hedged post-audit entries (`/handoff`, `/prototype`). (2) **Learnings-routing fix** (#25, closes #24): `SESSION_RUNNER.md` Phase 3C dual-audience recomposition — adopters record learnings in `CLAUDE.md` Adaptations, canonical repo appends to the seed table; false "table starts empty" caption corrected; `HOW_TO_USE.md` retargeted. (3) **CLAUDE.md size budget** (#27, closes #26): `BOOTSTRAP.md` Step 5 size-budget rule + extract-to-`PROJECT_LEARNINGS.md` overflow valve + plain-link (not `@`-import) pointer warning. (4) `tools/__pycache__/` gitignored (#23). No principle, phase, gate, workstream, or FM changes; FM count stays 25.
 - **v2.6.2:** Docs-only dot release resolving issue #22 (Pocock skill enablement gap, rmsharp investigation) under **strategy C (Reactive)**: enablement is per-request via adopter `CLAUDE.md` adaptations or `/grill-with-docs` lazy-creation — no methodology-shipped templates (preserves "methodology recommends; methodology does not reimplement"). (1) **Bounding axis:** new "Phases spanned (compression risk)" column on `RECOMMENDED_SKILLS.md`'s external-skills table — verbatim-verified per-skill phase spans at the pinned SHAs (`/grill-with-docs` + `/improve-codebase-architecture` = HIGH, Phase 2→3→5; `/diagnose` = medium by design); reading-guidance paragraph (plan the stop point before invoking high-compression skills; "A skill is not a phase" cross-ref); in-file forward-compat note to restate in vertical-slice vocabulary if #20/#21 land. (2) **`/triage` contradiction:** resolved by footnote — citation stands, adopters establish the label mapping themselves per `DEVELOPMENT_WORKSTREAM.md` §Issue Lifecycle instead of running the declined `/setup-matt-pocock-skills`. Issues #20/#21 remain open and unprejudged. No principle, phase, gate, workstream, or FM changes.
 - **v2.7:** Vertical-slice model — the first change to the unit of "1 and done" (issues #20/#21, adopted by operator decision 2026-06-12). One deliverable MAY be a **verified vertical slice** (one capability end to end) under four mandatory gates: (a) layer set pre-declared in a plan-mode contract before any code (prior-session approval satisfies; implementing session re-verifies at Orient), (b) checkpoint commit at every layer boundary (5-file cap unchanged, clarified *per-commit* in `SAFEGUARDS.md`), (c) full build/test matrix + exhaustive grep at *each* boundary, (d) per-surface faithful verification ("all tests ran" is not automatically faithful). **Evidence-gated, not self-certified** — missing per-boundary artifacts revert the session to horizontal scope at the last clean checkpoint. Recoverability, not verifiability, is the ceiling on slice size; seven boundary types stay non-collapsible (irreversible migrations, operator-approval gates, plan ↔ implementation, cross-toolchain cutovers, live-device verification, unfaithful-suite surfaces, recoverability ceiling). New `SESSION_RUNNER.md` §Vertical Slice Sessions; Principle 9 extended; guardrails: new **FM #26** "mega-session masquerading as a vertical slice" (appended — 1–25 unchanged) + degradation row, FM #17 anti-erosion clause (the allowance ADDS a gate, removes no step), FM #18 discriminator + degradation-row exception for pre-declared slices, 6-phase-model mirroring note. Touch-ups: `RECOMMENDED_SKILLS.md` phase-map guidance restated in vertical-slice vocabulary (v2.6.2's forward-compat note discharged); §Recommended Skills "compress two sessions into one" sentence reworked. Strict horizontal "1 and done" remains the default and the fallback; no phase, gate, or workstream changes.
+- **v2.7.2:** Docs-only dot release bundling three external contributions (rmsharp PRs #29/#30/#31), each resolving one PR-14-review issue. (1) **Agent-level memory generalization** (#29, closes #16): `starter-kit/CONTEXT_TEMPLATE.md`'s "CONTEXT.md vs auto-memory" guidance is renamed "CONTEXT.md vs **agent-level memory**" — the second persistence layer is now named by capability (persistence scoped to the agent across all projects) rather than by Claude Code's brand term, with a one-line gloss that agents implement it under different names (Claude Code's auto-memory, Cursor's memories, Cody's preferences); keeps the methodology agent-independent. (2) **Workstream skill-citation convention** (#30, closes #17): `starter-kit/RECOMMENDED_SKILLS.md` §How this index is used gains a **count-threshold rule** — one recommended skill is cited inline at the point of recommendation (e.g. `DEVELOPMENT_WORKSTREAM.md` §Issue Lifecycle → `/triage`; `ARCHITECTURE_WORKSTREAM.md` §Refactor Heuristics → `/improve-codebase-architecture`), while two or more skills for one workstream are promoted to a dedicated `## Recommended Skills` section with a purpose→skill table (e.g. `AUDIT_WORKSTREAM.md`'s `/code-review`/`/review`/`/security-review`). A single-row table is heavier than the inline citation it would replace, so the section earns its overhead only at ≥2 skills; `workstreams/TEMPLATE_WORKSTREAM.md` gains a matching optional "Recommended Skills" scaffold (delete-if-one, keep-if-≥2). (3) **Cross-reference completeness Learning** (#31, closes #19): the v2.6 self-review heuristic is formalized as **Learning #7** in the `SESSION_RUNNER.md` Learnings table (appended — 1–6 unchanged) plus a Phase 3F pre-commit hook — when a change adds a citation/index/glossary/cross-reference or modifies a numbered set (FMs, principles, phases, anti-patterns, learnings), `git diff` shows only the present-side edit, so grep each cited destination and each nearby set-size claim before committing. No principle, phase, gate, workstream, or FM changes; FM count stays 26.

README.md

@@ -231,6 +231,15 @@ Developed by Terrell Deppe (KJ5HST) using Claude Code (Anthropic) during develop
 
 The framework is agent-independent — it works with any AI coding agent that supports persistent files and session-based interaction. It also works for human developers, though the Session Runner and known failure modes are specifically tuned for AI agent tendencies.
 
+### What's New in v2.7.2
+
+Docs-only dot release bundling three contributor fixes from the PR-14 review backlog — issues [#16](https://github.com/KJ5HST/methodology/issues/16), [#17](https://github.com/KJ5HST/methodology/issues/17), and [#19](https://github.com/KJ5HST/methodology/issues/19). No principle, phase, gate, workstream, or failure-mode changes.
+
+- **Agent-level memory, named generically (#16).** `starter-kit/CONTEXT_TEMPLATE.md`'s "CONTEXT.md vs auto-memory" guidance becomes "CONTEXT.md vs **agent-level memory**" — the cross-project persistence layer is now named by what it does, not by Claude Code's brand term, with a gloss that other agents implement it under their own names (Cursor's memories, Cody's preferences). Keeps the methodology agent-independent.
+- **A convention for citing skills from workstreams (#17).** `RECOMMENDED_SKILLS.md` now states the rule the docs already followed: cite **one** recommended skill inline at the point of recommendation; promote **two or more** for the same workstream to a dedicated `## Recommended Skills` section with a purpose→skill table. `TEMPLATE_WORKSTREAM.md` ships a matching optional scaffold.
+- **Cross-reference completeness as Learning #7 (#19).** A new Learnings-table row plus a Phase 3F pre-commit hook codify the self-review check `git diff` can't do for you: when you add a citation or grow a numbered set (failure modes, principles, learnings…), the defect is the *missing* destination-side edit — so grep each cited destination and each nearby count claim before committing.
+- **Backward compatible.** Docs-only. `SESSION_RUNNER.md` reaches adopters via `bin/sync`; the framework/template docs (`CONTEXT_TEMPLATE.md`, `RECOMMENDED_SKILLS.md`, `TEMPLATE_WORKSTREAM.md`) are outside `bin/sync`'s 3-file set and update via the manual overlay or the "Update methodology using …" step.
+
 ### What's New in v2.7
 
 The first change to the unit of **"1 and done"** since the rule was written: one deliverable MAY now be a **verified vertical slice** — one capability end to end — under hard gates. Adopted from issues [#20](https://github.com/KJ5HST/methodology/issues/20) (model) and [#21](https://github.com/KJ5HST/methodology/issues/21) (guardrails).

starter-kit/SESSION_RUNNER.md

@@ -253,6 +253,7 @@ If you cannot runtime-verify (requires hardware, external service, CI), state th
 Before committing:
 
 - **Remove debug instrumentation added during this session.** Tagged debug logs (per `/diagnose` — see [`RECOMMENDED_SKILLS.md`](RECOMMENDED_SKILLS.md)) make this a single grep; ad-hoc prints make it manual. Either way, do not commit instrumentation that was meant to be temporary.
+- **Verify cross-references added or changed this session.** If you added a citation, index entry, glossary term, or cross-reference between files, or modified a numbered set (FMs, principles, phases, anti-patterns, learnings), grep each cited destination to confirm it resolves and grep nearby prose for set-size claims that may have drifted. `git diff` shows the present-side edit, not the missing destination (Learning #7).
 
 Then commit all changes with a descriptive message.
 
@@ -343,6 +344,7 @@ These are documented tendencies. The agent must actively guard against them.
 | 4 | **Verify a plan's output against its completion criteria — not against session duration or count.** Execution speed is not evidence of plan quality, and "fits in one session" is not a planning goal. A plan whose phases each fit cleanly in one session and produce work that matches the completion criteria is high-quality; a plan that tempts the executor to bundle phases or skip close-out to "finish faster" is not. The "1 and done" rule does not bend for high-quality plans. | Field observation (refined from rmsharp feedback on issue #7) | When evaluating a plan or judging plan quality. Resist the temptation to read execution speed as plan quality — they are uncorrelated when the protocol holds. |
 | 5 | Code review is a distinct deliverable, not overhead. Reviews that produce actionable plans (exact code snippets, line numbers, implementation order) have higher ROI than vague "this needs improvement" feedback. A review that doesn't identify specific changes a future session can execute is incomplete. | Field observation | When the session deliverable is a code review. Output an actionable plan, not a critique. |
 | 6 | A plan written from memory of a file read is an assumption-level claim. Reading implementations before estimating complexity catches wrong assumptions early — estimating from a backlog description alone is unreliable. This is a special case of FM #11 (gaps from memory) and FM #20 (edit from memory) applied to planning. | Field observation | Before estimating effort or scope for any phase, read the actual implementation file. The backlog description is a hint, not a spec. |
+| 7 | **Cross-reference completeness at self-review.** When a change adds a citation, an index entry, a glossary term, or any cross-reference between files — and especially when it modifies a numbered set (FMs, principles, phases, anti-patterns, learnings) — `git diff` confirms only the present-side edit; the defect is the *absence* on the destination side, which the diff cannot surface. Grep each cited destination to confirm the citation resolves, and grep nearby prose for set-size or named-membership assertions ("N failure modes", "the three workstreams") that may have drifted. Bidirectional: a new index entry needs its destination to cite back; a grown numbered set needs its size claims re-counted. | Three v2.6 instances: stale "24 failure modes" count, understated `/diagnose` + `/grill-with-docs` citation rows, aspirational `/init` citation (issue #19) | When a change touches any cross-file citation, index, glossary term, or numbered set. At self-review, grep destinations and size claims before committing — do not rely on `git diff` alone. |
 
 ---