release: v7.0.0 — voice-axis cutover, deepseek free tier, Polar payment route

.cursor/rules/patina.md

@@ -4,15 +4,15 @@
 
 Patina is a **Claude Code skill** that detects and removes AI writing patterns from Korean, English, Chinese, and Japanese text. It rewrites AI-sounding text into natural, human-like prose while preserving meaning through the Meaning Preservation System (MPS).
 
-The project uses a **plugin architecture**: patterns are plugins (`patterns/{lang}-*.md`), profiles are themes (`profiles/*.md`). Inspired by oh-my-zsh.
+The project uses a **plugin architecture**: pattern packs (`patterns/{lang}-*.md`), Document Type policies (`document-types/*.md`), and optional Persona v2 voice definitions (`personas/<lang>/*.md`).
 
 ## File Structure Conventions
 
 - **Pattern packs**: `patterns/{lang}-{category}.md` (e.g., `en-content.md`, `ko-language.md`)
   - Must have valid YAML frontmatter: `pack`, `language`, `name`, `version`, `patterns`
   - Pattern definitions use `### N. Pattern Name` headings
-- **Profiles**: `profiles/{name}.md` (e.g., `blog.md`, `academic.md`)
-  - Define `voice-overrides` and `pattern-overrides`
+- **Document Types**: `document-types/{name}.md` — genre, structure, and `pattern-overrides`
+- **Personas**: `personas/{lang}/{name}.md` — optional voice-only Persona v2 frontmatter
 - **Core definitions**: `core/voice.md`, `core/scoring.md`
 - **Examples**: `examples/{number}-success-01.md`, `{number}-failure-01.md`
   - English examples: `examples/en-{number}-success-01.md`
@@ -39,21 +39,19 @@ The project uses a **plugin architecture**: patterns are plugins (`patterns/{lan
 6. Update `README.md` pattern tables
 7. Bump the pack's `version:` frontmatter
 
-## When Adding a Profile
-
-1. Copy `profiles/default.md` as a template
-2. Define `voice-overrides` (amplify/allow/suppress per voice dimension)
-3. Define `pattern-overrides` (amplify/normal/reduce/suppress per pattern number)
-   - Use language-scoped overrides to avoid cross-language number collisions:
-     ```yaml
-     pattern-overrides:
-       ko:
-         8: amplify
-       en:
-         8: amplify
-     ```
-4. If the profile needs custom AI/fidelity balance, add to `.patina.default.yaml` under `scoring.combined-weights`
-5. Update `README.md` profile table
+## When Adding a Document Type
+
+1. Copy `document-types/default.md` as a template.
+2. Set `document-type:` to the filename stem.
+3. Define `purpose`, `audience`, `structure`, `style`, `avoid`, and
+   language-scoped `pattern-overrides` in frontmatter.
+4. Keep Persona voice, Register, and verification fields out of the policy.
+5. Add a matching key under `.patina.default.yaml`
+   `scoring.combined-weights.document-type` only when scoring evidence justifies
+   a non-default balance.
+
+Create reusable voice through Persona v2 (`patina persona new`). Persona schema
+validation rejects Document Type, Register, pattern-policy, and safety fields.
 
 ## When Modifying SKILL.md
 
@@ -68,14 +66,13 @@ The project uses a **plugin architecture**: patterns are plugins (`patterns/{lan
 
 ## Version Management
 
-- `.patina.default.yaml` `version:` is the **source of truth**
-- `SKILL.md` and `README.md` must match it
-- Pattern pack versions are independent — bump when patterns change
-- Profile versions are independent
+- `package.json` `version` is the **source of truth**
+- `SKILL.md`, `.patina.default.yaml`, and `README.md` must match it
+- Pattern pack and Document Type versions are independent
 
 ## Language Synchronization Rule
 
-All 4 languages (ko, en, zh, ja) must maintain **29 patterns each** (116 total). When adding a universal pattern:
+All 4 languages (ko, en, zh, ja) currently maintain **46 patterns each** (184 total). When adding a universal pattern:
 - Same pattern number across all 4 languages
 - Same semantic category
 - Language-specific watch words and examples

.patina.default.yaml

@@ -4,41 +4,26 @@ language: ko              # Korean (default) -- auto-loads all ko-*.md patterns
                           # language: zh      -- Chinese -- auto-loads all zh-*.md patterns
                           # language: ja      -- Japanese -- auto-loads all ja-*.md patterns
                           # Override with CLI: --lang en / --lang zh / --lang ja
-profile: default
-output: rewrite           # rewrite | diff | audit | score
+document-type: default     # pattern policy: default, blog, academic, technical,
+                           # formal, social, email, legal, medical, marketing,
+                           # narrative, instructional, casual-conversation,
+                           # code-comment, commit-message, release-notes, namuwiki
+# `formal` means structured documents; `casual-conversation` means chat/message
+# structure. Neither selects casual/professional delivery — use `register`.
+output: rewrite            # rewrite | diff | audit | score
 
-
-# Tone categorization (v3.10): register axis (casual/professional) for ko/en.
-# Resolution: explicit --tone > config tone > config profile.
-# auto = heuristic single-tone detection at SKILL.md Phase 4.5b.
-# Unset = profile-only legacy behavior (regression-safe default).
-tone:                       # one of: casual, professional, auto (register only)
-                            # Override with CLI: --tone <name>
-                            # zh/ja with explicit tone: warns and falls back to profile.
-
-# Persona harness (v1): Korean rewrite defaults to the conservative preserve
-# persona. Non-rewrite and non-Korean surfaces do not apply persona implicitly.
-persona: preserve
+# Optional rewrite axes. Omit persona to preserve source voice; omit register
+# to preserve the source register.
+register:                  # casual | professional
+# --persona applies a reusable voice fingerprint from personas/{lang}/.
+persona:
 personas:
-  library: personas/ko
   thresholds:
-    schema: patina.persona.thresholds.v1
-    # source stays `placeholder` because the advisory tuning below is informed by
-    # observation, not a formal 2-round promotion. The ENFORCING floors, however,
-    # are the product's core safety contract and do not depend on calibration.
+    # Persona match and surface churn are advisory voice-quality signals.
+    # Meaning, numbers, and fidelity are enforced by the global rewrite and
+    # --verify paths whether or not a persona is active.
     source: placeholder
-    calibration_rounds_required: 2
-    # --- ENFORCING (safety gate; a miss sets a non-zero exit, non-destructive) ---
-    # Meaning + facts only. Enforced when evaluated (mps/fidelity need --verify or a
-    # backend-reported score); dropped source numbers always enforce deterministically.
-    mps_floor: 70
-    fidelity_floor: 70
-    # --- ADVISORY (warn only; never blocks or changes the exit code) ---
-    # persona_match: voice-quality signal. Observed real KO persona-match runs ~35-56,
-    # so 45 flags the low end without firing on every rewrite.
     persona_match_min: 45
-    # churn: surface token change, NOT a meaning signal. Live KO humanizing rewrites
-    # churn ~0.5-0.85 while preserving meaning, so 0.9 flags only near-total rewrites.
     churn_max: 0.9
 
 # Supported languages: ko, en, zh, ja

CHANGELOG.md

@@ -12,23 +12,7 @@ All notable changes to patina. Dates are release dates (YYYY-MM-DD).
 Semver rationale: patch | minor | major — explain whether this changes patterns, schemas, CLI behavior, or docs only.
 ```
 
-## Unreleased
-
-### Changed
-
-- The v6.4 preflight hold now certifies the Polar payment route instead of the
-  declined Lemon Squeezy account: the checkout evidence binding table carries
-  exactly the Polar production tuple (`PAY-B-20260729-POLAR-ea8385dc-4c9c3f17`,
-  `https://buy.polar.sh/polar_cl_…`), the hold validates the
-  `PAY-B-BINDING-POLAR-v1` and `PAY-LIVE-RUNTIME-POLAR-v1` artifacts,
-  `POLAR_APPROVAL` replaces `LS_APPROVAL`, and the secret-manager blocker names
-  `PATINA_LICENSE_PROVIDER`/`POLAR_ORGANIZATION_ID`/`POLAR_PRO_BENEFIT_ID`. The
-  retired staging chain is recorded as superseded by the production zero-amount
-  purchase evidence, not silently dropped; Lemon Squeezy records stay on disk,
-  hash-frozen, as history. Checkout remains disabled pending the owner's
-  env-side enable sequence.
-
-## 7.0.0 — 2026-07-30
+## 7.0.0 — 2026-08-04
 
 **Removes the retired iterative rewrite product contract, gives scoring and verification settings neutral ownership, and preserves the comparator only as unsupported research.**
 
@@ -39,23 +23,35 @@ Semver rationale: major — removes public configuration keys and the obsolete a
 - The retired iterative mode is gone from the agent skill, CLI migration tombstone, default configuration, help/support surfaces, generated API, and current documentation. Node users keep `--verify`; agent-skill users keep the independent `/patina --strict` flow.
 - Loop-only `enabled`, `target-score`, `max-iterations`, and `plateau-threshold` settings are no longer product configuration.
 - The unused top-level `structured-output` setting is deleted. The scorer's explicit runtime `responseFormat` API remains available to supported callers.
+- **`--profile`, `--tone`, and `--formality` are retired** along with the `profiles/` and `examples/tones` catalogs; the CLI rejects them with migration errors pointing at the replacement axes below. The `preserve` persona files are gone — omitting `--persona` now preserves the source voice by construction.
 
 ### Changed (breaking)
 
 - Move `category-weights`, `combined-weights`, and `severity-points` from the former top-level loop-settings block to `scoring`.
 - Move the shared `mps-floor` and `fidelity-floor` values to `verification`. Their defaults remain 70/70.
 - Keep `personas.thresholds.mps_floor` and `personas.thresholds.fidelity_floor` separate; persona gate overrides do not affect `--verify`, and verification overrides do not affect persona gates.
 - Rename the opt-in A/B arm to `iterative-baseline` and move its runner under `scripts/`. It remains packaged for `npm run quality:rewrite-ab`, but is unsupported research and is absent from CLI/help/config, hosted APIs, and generated public API docs.
-- Persona, profile, and tone behavior is unchanged. Persona remains the reusable voice axis, tone remains the KO/EN register override, and profile-policy consolidation remains follow-up work.
+- The rewrite-axis model is rebuilt (see Added): Document Type owns document policy, Persona v2 stays the optional reusable voice axis, and Register owns delivery. Meaning preservation and verification remain global; no axis can weaken them.
 
 ### Added
 
 - `patina --score --offline` runs the deterministic scoring layer without resolving a backend or credential. It supports markdown/JSON output and `--exit-on`; LLM-judged categories are marked unavailable, and a missing numeric local score fails instead of silently passing.
+- Three independent rewrite axes replace the overloaded profile/tone/formality model: 17 Document Types own document policy, 12 language-scoped Persona v2 files optionally supply reusable voice, and Register independently selects `casual` or `professional` delivery. Omitted Persona and Register preserve the source.
+- Document Type files now use a validated structured frontmatter contract (`purpose`, `audience`, `structure`, `style`, `avoid`, language-scoped `pattern-overrides`). Invalid or cross-axis custom policies fail before prompt assembly.
+- Rewrite prompts state axis ownership and precedence explicitly. Document Type cannot manufacture voice, Persona cannot choose document policy or Register, and Register cannot alter structure or claims.
+- The example catalog now includes comparable KO/EN rewrite-axis fixtures for academic, instructional, marketing, narrative, casual, and professional behavior.
+- Server-paid free-tier rewrites on DeepSeek get a measured reasoning cut (`reasoning_effort: low`, ~44s instead of ~90s at gate-parity quality), env-tunable via `PATINA_FREE_REWRITE_REASONING=low|medium|high|off`; the MPS/fidelity judges carry the same cut on gemini and deepseek. BYOK and Pro rewrites never receive the field. The streaming client gains `extraBody` passthrough with protocol fields protected.
 
 ### Fixed
 
 - Rewrite stdout now contains prose only. Plain, fenced, and blockquoted tone footers are stripped, while `--format json` retains structured tone metadata, including the model-resolved value from `--tone auto`.
 - The agent skill now uses the CLI's config precedence and canonical project filename: installed `.patina.default.yaml` → `~/.patina.yaml` → project `./.patina.yaml` → command arguments.
+- A model cached as rejecting `temperature` can no longer recover the field from `extraBody` in either LLM client.
+- The BYOK pricing card no longer claims the key "stays in your browser": it is sent only with your requests, relayed by the server, and never stored or logged.
+
+### Operations
+
+- The v6.4 preflight hold now certifies the Polar payment route instead of the declined Lemon Squeezy account: the binding table carries exactly the Polar production tuple (`PAY-B-20260729-POLAR-ea8385dc-4c9c3f17`), `POLAR_APPROVAL` replaces `LS_APPROVAL`, and the retired staging chain is recorded as superseded by production zero-amount purchase evidence. Checkout was opened by the owner on 2026-08-04 after the full gate sequence; the hosted free tier now serves on `deepseek-v4-flash`.
 
 ### Migration
 

CONTRIBUTING.md

@@ -1,10 +1,10 @@
 # Contributing to Patina
 
-Thanks for considering a contribution. Patina is a pattern-based tool, so the most impactful contributions are usually new patterns, better examples, or profile refinements.
+Thanks for considering a contribution. Patina is a pattern-based tool, so the most impactful contributions are usually new patterns, better examples, Document Type policies, or Persona refinements.
 
 ## Public vs. Internal Docs
 
-User-facing documentation lives in `README*.md`, `docs/`, `examples/`, `patterns/`, `profiles/`, and the skill entrypoints. Maintainer or agent notes (backlogs, runbooks, launch playbooks) are kept out of this repository entirely — they live in the maintainer's private workspace, and `docs/internal/` is gitignored as a guard. Only real user-facing facts get promoted into the public docs list.
+User-facing documentation lives in `README*.md`, `docs/`, `examples/`, `patterns/`, `document-types/`, `personas/`, and the skill entrypoints. Maintainer or agent notes (backlogs, runbooks, launch playbooks) are kept out of this repository entirely — they live in the maintainer's private workspace, and `docs/internal/` is gitignored as a guard. Only real user-facing facts get promoted into the public docs list.
 
 When moving a root-level Markdown file, link it from `README.md` if it is public; if it is maintainer-internal, it does not belong in the repository.
 
@@ -73,7 +73,7 @@ False positives are expected, especially for academic, encyclopedic, legal, corp
 1. Use the false-positive issue template.
 2. Include language, genre/register, score/audit excerpt, and the specific pattern that over-fired.
 3. Remove private text or replace it with a minimal redistributable excerpt.
-4. Suggest whether the fix should be an exclusion rule, lower severity, profile override, or benchmark fixture.
+4. Suggest whether the fix should be an exclusion rule, lower severity, Document Type `pattern-overrides` change, or benchmark fixture.
 
 Maintainers should prefer tightening exclusions over deleting patterns outright.
 
@@ -160,11 +160,18 @@ See [docs/HARNESS.md](docs/HARNESS.md) for the full measurement-tool map.
 - Add a target-language false-positive note when a phrase is normal in that register.
 - Keep examples redistributable; do not paste private user text.
 
-## Adding a Profile
+## Adding a Document Type
 
-Profiles live in `profiles/{name}.md`. Copy an existing one (e.g. `blog.md`) and adjust:
-- `voice-overrides`: which voice dimensions to amplify/suppress
-- `pattern-overrides`: per-language, per-pattern-id actions. `suppress` is applied deterministically (the pattern is dropped from the prompt for that language); `reduce`/`amplify` are advisory and reinforced only by the prose body.
+Document Types live in `document-types/{name}.md`. Copy an existing policy
+(for example `blog.md`), set `document-type:` to the filename stem, and define
+`scope`, `purpose`, `audience`, `structure`, `style`, `avoid`, and
+language-scoped `pattern-overrides`. `suppress` is applied deterministically;
+`reduce`/`amplify` document policy intent but do not currently change a runtime
+weight. Keep Persona voice, casual/professional Register markers, and
+verification thresholds out of this axis.
+
+For reusable voice, create a Persona v2 with `patina persona new`. Persona
+frontmatter is validated to reject document policy, Register, and safety fields.
 
 ## Pattern Staleness
 
@@ -176,14 +183,14 @@ How we handle this:
 - **Quarterly review:** Maintainers follow [`process/pattern-freshness.md`](process/pattern-freshness.md) for corpus freeze windows, promotion thresholds, and frontmatter metadata
 - **Lexicon provenance:** Newly mined or re-mined lexicon entries need `added`, `source`, and `last_validated` provenance before changing shipped behavior; run `npm run lexicon:freshness` to verify sidecars match the shipped entries
 - **Version notes:** Each pattern pack has a `version` field — bump it when patterns change
-- **No deletion without replacement:** We don't remove patterns outright; we mark them as `low` severity or move them to `reduce` in profiles
+- **No deletion without replacement:** We don't remove patterns outright; we mark them as `low` severity or move them to `reduce` in a Document Type policy
 
 ## Versioning Policy
 
 Patina uses semantic versioning for both CLI behavior and pattern-pack compatibility.
 
 - **Major:** remove or renumber patterns, break config/result schemas, change public CLI semantics, or make existing pattern packs incompatible.
-- **Minor:** add a pattern, language, profile, mode, backend, benchmark schema field, or contributor-facing workflow.
+- **Minor:** add a pattern, language, Document Type, Persona, mode, backend, benchmark schema field, or contributor-facing workflow.
 - **Patch:** fix bugs, adjust severity/exclusions, clarify examples, update docs, refresh benchmark fixtures without changing schemas.
 
 Every changelog entry should include a short semver rationale line so downstream users know whether to pin, test, or upgrade normally.