docs: Phase K — CHANGELOG 1.3.0, README profiles section, LESSON 0.9

plugins/preview-forge/CHANGELOG.md: full 1.3.0 entry documenting profiles,
Rule 9 drift detector, P0-B cost regression, surface detection, preview
cache, all 11 agent budget changes, auto-extend SCC, tools-vs-agents split.

README.md: new "Profiles (v1.3+)" section with comparison table and
"Cost regression + drift detection" subsection. Quick Install snippet now
shows --profile=standard|pro|max usage examples.

memory/LESSONS.md: LESSON 0.9 — "한 flag 매트릭스 대신 profile 단일화".
Documents the design decision to collapse 4 boolean flags × 16 combos
into a single --profile flag. References devops-architect panel vote.

No version-number edits — release-please will handle the bump on merge
via .release-please-manifest.json and extra-files jsonpaths.

verify-plugin.sh: 45/45 local checks pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: ComBba

README.md

@@ -64,10 +64,32 @@ single HTML file, print-friendly.
 # 4. Initialize memory (first time only)
 /pf:bootstrap
 
-# 5. Run
+# 5. Run (profile defaults to `pro`)
 /pf:new "한 줄 아이디어"
+
+# …or pick a profile explicitly:
+/pf:new "demo-class idea"     --profile=standard   # ~60k tok · 2×5 eng · 9 previews
+/pf:new "real project"        --profile=pro         # ~250k tok · 3×5 eng · 18 previews  (default)
+/pf:new "production launch"   --profile=max         # ~600k tok · 5×5 eng · 26 previews
 ```
 
+## Profiles (v1.3+)
+
+| Profile | Previews | Eng teams | Panels | SCC iter | P95 ceiling | Use for |
+|---|---|---|---|---|---|---|
+| **standard** | 9 | 2×5 (BE+FE) | keyword-trigger | 3 | ~60k tok / 25 min | Demo · prototyping |
+| **pro** *(default)* | 18 | 3×5 (+DB) | keyword-trigger + escalation | 4 | ~250k tok / 70 min | Real projects |
+| **max** | 26 | 5×5 (all) | always-on | 5 | ~600k tok / 160 min | Production · baselines |
+
+- `--previews=N` overrides the count (bounded by `max_user_expand` = 26).
+- `--no-cache` bypasses the PreviewDD-level cache (default TTL: 7 days for standard/pro, never cached for max).
+- Full spec: [`plugins/preview-forge/profiles/`](plugins/preview-forge/profiles/).
+
+### Cost regression + drift detection (v1.3+)
+
+- **Rule 9 idea-drift detector** (`hooks/idea-drift-detector.py`) catches the failure where Gate H1 picks product A but SpecDD/Engineering drift to product B. Containment coefficient over token sets (no external ML deps). Block threshold 0.3, warn at 0.4.
+- **P0-B cost-regression sentinel** (`hooks/cost-regression.py`) compares `cost-snapshot.json` against the active profile's P95/hard ceiling every 30s. Hard breach triggers auto-pause + AskUserQuestion handoff.
+
 ## Updating
 
 We release patches and feature updates frequently (see

plugins/preview-forge/CHANGELOG.md

@@ -2,8 +2,132 @@
 
 All notable changes to this plugin will be documented in this file.
 
-## [1.0.0] — 2026-04-22 (in progress)
+## [1.3.0] — 2026-04-23 (in progress)
+
+### Added — `--profile` system (standard · pro · max)
+
+Single flag replaces the old `--lean/--previews=N/--single-team/--skip-panels`
+matrix (4 booleans × 16 combinations) that the devops-architect panel
+vote rejected. `--profile=pro` is the default.
+
+- **`profiles/standard.json`**: 9 previews · 2×5 eng (BE+FE) · keyword-triggered
+  panels · SCC iter=3 · P95 60k tok / 25 min. Demo / prototyping.
+- **`profiles/pro.json`**: 18 previews · 3×5 eng (+DB) · keyword-triggered with
+  escalation · SCC iter=4 · P95 250k tok / 70 min. **Default**. Real projects.
+- **`profiles/max.json`**: 26 previews · 5×5 eng (all disciplines) · all panels
+  always-on · SCC iter=5 · P95 600k tok / 160 min. Production / baselines.
+- **`schemas/pf-profile.schema.json`**: Draft-07 schema, CI-validated.
+
+### Added — Layer-0 Rule 9 (idea-drift detector)
+
+Catches the failure mode "Gate H1 picks P10 API but SpecDD/Engineering
+drift to P02 Slack" — usually template caching or agent memory leak.
+
+- **`hooks/idea-drift-detector.py`**: PreToolUse hook. Containment coefficient
+  `|chosen ∩ incoming| / |chosen|` over stopword-filtered tokens. Default
+  threshold 0.4. No external deps (stdlib only, LESSON 0.4).
+- **Scope**: `runs/*/specs/SPEC.md` · `openapi.yaml(.lock)?` · `apps/*/README.md`.
+- **Exit**: 0 allow · 1 warn (0.3–0.4) · 2 block (<0.3).
+- **Bypass**: `PF_DRIFT_BYPASS=1` + `PF_DRIFT_REASON=...` for intentional scope expansion.
+
+### Added — P0-B cost-regression sentinel
+
+- **`hooks/cost-regression.py`**: standalone CLI + monitor watcher. Reads
+  `runs/<id>/cost-snapshot.json`, compares against profile's `cost_ceiling`
+  (P95 + hard). Writes blackboard row at severity `warn` or `alert`.
+- **`monitors/monitors.json`**: cost-regression watcher runs every 30s
+  per run dir.
+- Hard-ceiling breach triggers M1 Run Supervisor to pause + AskUserQuestion.
+
+### Added — Surface-type detection (Proposal #2)
+
+- **`scripts/detect-surface.sh`**: regex-gate over idea.json → classifies
+  rest-first / ui-first / hybrid. Bilingual keyword banks (EN + KO).
+  Outputs `{"surface": ..., "stack_hint": ...}` JSON.
+- Fixes the "Next.js 16 blindly applied to API-first products" failure
+  reported by system-architect panel (Minutes.ai case).
+- Stack routing: rest-first → nestia · ui-first → Next.js 16 · hybrid → both.
+
+### Added — PreviewDD cache (Proposal #11)
+
+- **`scripts/preview-cache.sh`** {key|get|put|invalidate|prune}.
+- Cache key: `sha256(idea + advocate_set + model_version + profile)`.
+- TTL from profile: standard/pro 7 days, max never (0 = production safety).
+- `/pf:new --no-cache` forces bypass.
+- ~60% token saving on demo re-runs with identical idea.
+
+### Changed
+
+- **Agent budgets** are now profile-aware across all tiers:
+  - Preview Advocates (26): 20K → profile (12K / 14K / 20K)
+  - Panel Leads (4): 120K → profile (60K / 84K / 120K)
+  - Panel Members (40): 40K → profile (24K / 28K / 40K)
+  - QA Leads (4): 80K → profile (48K / 56K / 80K)
+  - QA Members (14): 40K → profile (24K / 28K / 40K)
+  - SCC Lead: 80K → profile (56K / 64K / 80K)
+  - M3 Chief Engineer: 120K → profile (84K / 100K / 120K)
+- **`/pf:new`** accepts `--profile=standard|pro|max` · `--previews=N` ·
+  `--no-cache`. Default = `pro`.
+- **`/pf:status`** + **`/pf:budget`** now show active profile + budget-to-ceiling.
+- **SCC auto-extend**: if error count is decreasing (`e_i < e_{i-1}`),
+  allow up to +2 iter beyond `max_iter` for convergence.
+- **Panel activation**: standard/pro `keyword-trigger` mode — panels only
+  activate if idea keywords match trigger list. pro auto-escalates to full
+  panel on unknown-unknown detection (advocate dispersion > 0.7).
+
+### Methodology
+
+- **`methodology/global.md`** gains Rule 9 (idea-drift fidelity).
+- **QA tools-vs-agents separation** formalized in `qa/security/secqa-lead.md`:
+  tools (semgrep, gitleaks, axe-core, lighthouse, owasp-zap) always run
+  regardless of agent count. Only agents scale with profile.
+
+### CI
+
+- Validate all 3 profiles against schema on every PR.
+- New test matrices: idea-drift (5 cases), cost-regression (6 cases),
+  detect-surface (3 cases), preview-cache (4 cases).
+- verify-plugin.sh now runs 45 checks (was 38).
+
+### Rationale
+
+All decisions ratified by 5-expert panel: system-architect, refactoring-expert,
+root-cause-analyst, quality-engineer, devops-architect. See PR body for the
+full debate summary.
+
+## [1.2.1] — 2026-04-22
+
+### Fixed
+- release-please manifest mode so plugin.json + marketplace.json sync on tag.
+
+## [1.2.0] — 2026-04-22
+
+### Added — Layer-0 Rule 8 (run artifact single-writer)
+
+- `hooks/factory-policy.py` gains run-artifact enforcement: only the M1
+  Run Supervisor can write to `runs/*/chosen_preview.json(.lock)?` ·
+  `.panel-recommended.json` · `design-approved.json(.lock)?` ·
+  `mitigations.json` · `panels/meta-tally.json` · `score/report.json` ·
+  `.frozen-hash`.
+- External out-of-band editors (sibling skills, other assistant sessions,
+  user manual edits) blocked with exit 2.
+- Bypass via `PF_WRITER_ROLE=supervisor` env var (supervisor slash-command flow).
+
+## [1.1.0] — 2026-04-22
+
+### Fixed — UX gap from first real run
+
+- **Gate H1 is now preview-selection + design-tweak (was design-only)**
+- `commands/design.md` rewritten: AskUserQuestion 4옵션 (Recommended /
+  Alt A / Alt B / Gallery of 26).
+- `chosen_preview.panel-recommended.json` backup of panel's original pick.
+
+## [1.0.0] — 2026-04-22
 
 ### Added
 - Plugin manifest (`.claude-plugin/plugin.json`)
 - Apache-2.0 license + NOTICE
+- 143 Opus 4.7 agents across 6 tiers
+- 14 slash commands (`/pf:*`)
+- 3 hooks (factory-policy, askuser-enforcement, auto-retro-trigger)
+- 3-DD methodology (PreviewDD → SpecDD → TestDD)

plugins/preview-forge/memory/LESSONS.md

@@ -9,6 +9,16 @@
 
 ## 0. 플러그인 개발 자체에서 배운 것 (bootstrap)
 
+### 0.9 한 flag 매트릭스 대신 profile 단일화 — 구성 표면적 최소화 (category 6 Plugin 배포, UX)
+
+- **문제**: v1.2.x의 e2e run이 2시간+ · 토큰 수백만 소모. "lean mode"로 축소하려는 v1.3.0 초안이 4개 boolean 플래그 (`--lean --previews=N --single-team --skip-panels`) → 16개 조합을 만들어 사용자가 "어떤 조합이 올바른 조합인지" 판단 불가. devops-architect 패널이 "이 매트릭스는 문서화도 테스트도 지원 불가"라고 거부.
+
+- **원인**: 세 가지 직교 축(속도·깊이·안전)을 독립 플래그로 노출하면 matrix explosion. 사용자는 의미 있는 '세트'를 원하지 축별 토글을 원하지 않음.
+
+- **해결**: v1.3.0에서 `--profile=standard|pro|max` 단일 플래그로 치환. 각 profile이 previews 수·eng 팀 수·panel 모드·SCC iter·budget ceiling을 통째로 묶음. `standard=demo/prototype`, `pro=default/real project`, `max=production/baseline`. 플래그 매트릭스 폐기. settings.json의 `defaultProfile`로 tenant/팀별 기본값 설정 가능. 개별 축 override는 `--previews=N` 같은 명시적 escape 플래그로만 제한.
+
+- **참조**: `plugins/preview-forge/profiles/{standard,pro,max}.json`, `schemas/pf-profile.schema.json`, v1.3.0 PR body의 5-전문가 패널 토론 요약, devops-architect 투표 ("config in settings.json not env, profile not flag-matrix").
+
 ### 0.8 Live run artifact에 외부 writer 금지 — single-writer 원칙 (category 9 Agent communication, 경쟁 조건)
 
 - **문제**: 2026-04-22 r-20260422-184337 실행 중, 외부 대화 세션(보조 assistant)이 `/pf:design` Gate H1이 열리기 전 `chosen_preview.json`을 P02 → P19로 직접 덮어썼음 (`blackboard.db` 11:19:15 user-override 이벤트). 같은 시점에 사용자의 플러그인 세션이 정식 Gate H1 AskUserQuestion을 실행 중이었고, 사용자는 P10(TP 단독 1위 API-first)을 선택. run-supervisor가 11:42:30에 chosen_preview를 P10으로 재작성, lock도 `1d8f9193…afbb`로 재계산. 결과적으로 P19 편집은 "stale override"로 `selection_metadata.prior_stale_override_noted`에 기록되고 폐기됨. 플러그인 자체는 **정확히 작동** — 사용자의 in-flow 선택이 외부 out-of-band 편집을 이기도록 설계됨. 하지만 혼란과 낭비 발생.