docs: mark context-budget campaign practically complete

README.md

@@ -103,6 +103,20 @@ The repository is designed so a fresh clone is publishable without further redac
 
 All in-repo examples use placeholders only: `<your-org>/<your-repo>`, `<workspace>`, `<project-root>`, `<your-api-key>`, `<contributor>`. Full publication discipline: [`docs/PUBLICATION_CHECKLIST.md`](docs/PUBLICATION_CHECKLIST.md). Defence-in-depth detail: [`docs/SECURITY.md`](docs/SECURITY.md).
 
+## Context budget
+
+The repo is optimized around a compact always-loaded kernel plus lazy-loaded depth:
+
+- [`CLAUDE.md`](CLAUDE.md) is the always-loaded operating kernel.
+- [`AGENTS.md`](AGENTS.md) and [`WARP.md`](WARP.md) are public-safe pointer files.
+- Commands, skills, agents, domains, recipes, and references are lazy-loaded only after routing.
+- Bulky examples, templates, deep workflows, and implementation detail live in local `references/` docs next to each skill or agent.
+- Runtime behavior is preserved; what's lazy stays out of context until needed.
+
+Reference extraction is now the standard pattern for large skills and agents: the operating contract (frontmatter, when-to-use, decision logic, safety rules, validation gates, output expectations) stays inline; everything else moves into `references/`. Measure with `bash scripts/context-budget-report.sh`. Policy: [`docs/CONTEXT_BUDGET.md`](docs/CONTEXT_BUDGET.md). Campaign log: [`docs/CONTEXT_BUDGET_PLAN.md`](docs/CONTEXT_BUDGET_PLAN.md).
+
+**Current status: the reference-extraction campaign is practically complete.** Large SKILL.md and high-value agent files were compacted into operating contracts; detailed examples, troubleshooting, and playbooks moved into lazy-loaded references. Remaining large files are mostly routing indexes, command bodies, vendored upstream references, or intentionally lazy material — further work is optional polish.
+
 ## Validation
 
 ```bash
@@ -194,6 +208,7 @@ If you fork this repo for personal use, the first commit on your fork should NOT
 | Adoption guide | [`docs/ADOPTION_GUIDE.md`](docs/ADOPTION_GUIDE.md) |
 | Compatibility matrix | [`docs/COMPATIBILITY.md`](docs/COMPATIBILITY.md) |
 | Context budget guidance | [`docs/CONTEXT_BUDGET.md`](docs/CONTEXT_BUDGET.md) |
+| Context budget — extraction campaign log | [`docs/CONTEXT_BUDGET_PLAN.md`](docs/CONTEXT_BUDGET_PLAN.md) |
 | Common failure modes | [`docs/TROUBLESHOOTING.md`](docs/TROUBLESHOOTING.md) |
 | Pre-publication safety checklist | [`docs/PUBLICATION_CHECKLIST.md`](docs/PUBLICATION_CHECKLIST.md) |
 | Hooks reference (all 16) | [`docs/HOOKS.md`](docs/HOOKS.md) |

decisions/2026-05-05-history-rewrite-2-pii-and-cross-repo.md

@@ -10,7 +10,7 @@ Operator invoked `/ultraplan` security audit ("ensure no leak whatsoever") on 20
 
 1. **Operator OS username** (`<scrubbed>`) embedded in 16 tracked files: 2 audit-doc absolute paths + 14 yaml `author:` fields.
 2. **Second-private-project file paths** (`~/<private-project>/...`) leaked across 25 lines in `docs/SECURITY_PLAYBOOK.md` — same exposure class as the 2026-05-03 incident.
-3. **Operator's real email** (`<scrubbed>@​gmail.​com`) embedded in the **author metadata of every commit** (32+ commits) — discovered during the all-history grep. NOT visible to forward-state file scans.
+3. **Operator's real email** (`<redacted-email>`) embedded in the **author metadata of every commit** (32+ commits) — discovered during the all-history grep. NOT visible to forward-state file scans.
 4. **Operator's full name** (`the operator`) still present in older commits' diffs as deletion lines from prior PII sweeps (29 blobs + 2 commit messages).
 
 Forward-state fix landed in `944f9fe` (now orphaned). This decision covers the destructive history rewrite that scrubs all four classes from every commit.
@@ -25,7 +25,7 @@ Single-pass `git filter-repo` rewrite with:
    - operator full name → `Repository owner`
    - operator OS username → `redacted-username`
    - operator email-prefix → `redacted-email`
-   - any `@​gmail.​com` → `@example.invalid` (catch-all)
+   - any personal email address → `@example.invalid` (catch-all)
    - `~/<private-project>` → `~/your-project`
 3. **`--replace-message`** with the same pattern set so commit-message bodies are scrubbed.
 4. **Force-push** to `origin/main`.
@@ -40,7 +40,7 @@ Fresh `git clone` verification (post-push, run on 2026-05-05 22:03):
 | Full name | 0 | 0 | 0 |
 | OS username | 0 | 0 | 0 |
 | Email-prefix | 0 | 0 | 0 |
-| `@​gmail.​com` | 0 | 0 | 0 |
+| Personal email pattern | 0 | 0 | 0 |
 | `~/<private-project>` | 0 | 0 | 0 |
 | Public handle (allowed) | 68 | (LICENSE / README / ELITE-OPS docs / decisions logs) |
 

docs/CONTEXT_BUDGET_PLAN.md

@@ -67,6 +67,41 @@ inflection point when no SKILL.md was the top context risk.
   inventory script counts any `.md` under `agents/` as an agent definition.
 - Preserve upstream vendored content (e.g., `skills/n8n/czlonkowski/*`).
 
+## Final campaign status
+
+Status: **practically complete (~95% by impact).**
+
+The campaign achieved its primary goal: high-frequency operating files
+were compacted, bulky examples and implementation detail were moved into
+lazy-loaded references, and runtime behavior was preserved across every
+extraction.
+
+End-state evidence:
+
+- No SKILL.md exceeds ~2,800 estimated tokens (top remaining: `react-bits` at ~2,765).
+- `pm-agent.md` dropped from ~3,459 → ~2,302 tokens (no longer in P1–P3 risk list).
+- `CLAUDE.md` compacted from ~2,285 → ~2,103 tokens; still within the 5,000-token budget threshold.
+- Top context risks are now `ALWAYS_LOAD_RISK` (the kernel, by design),
+  `ROUTING_INDEX_RISK` (lazy on dispatch), and `LAZY_REFERENCE_HEAVY`
+  (lazy on invocation).
+- All validation gates pass: `validate.sh`, `check-public-safety.sh`,
+  `audit-public-readiness.sh --quick`, gitleaks, trivy.
+
+Remaining large files are acceptable because they are primarily:
+
+- **Routing indexes** (`agents/REGISTRY.md`, domain `DOMAIN.md` files) — loaded only when routing fires.
+- **Lazy command bodies** (`commands/sc/*`, `commands/bmad/*`) — loaded only on command invocation.
+- **Lazy reference files** (per-skill `references/`, large `*.md` resources) — loaded only when the skill is selected.
+- **Vendored upstream material** (`skills/n8n/czlonkowski/*`) — intentionally left untouched to preserve upstream sync.
+
+Future work is optional polish unless a file becomes always-loaded,
+exceeds budget, or starts appearing repeatedly in high-priority
+context-risk reports. Re-open the campaign only if:
+
+1. A new SKILL.md exceeds 2,800 tokens.
+2. `CLAUDE.md` exceeds 4,000 tokens.
+3. A validation gate starts failing on count-drift or public-safety.
+
 ## Validator constraints that shape CLAUDE.md
 
 `scripts/validate.sh` requires the following strings to exist verbatim in