generate-images

name	generate-images
description	use_when: | Hero image üret, blog görseli oluştur, nano-banana image gen, 8K ultra realistic image, hero image regenerate. Kullanıcı "hero image üret", "blog görseli oluştur", "nano-banana image gen", "8K ultra realistic image" der ya da `/pseo-generate-images` çağırır. master.xlsx[new_content_plan].image_prompt + alt_text source consume; project-config[content_settings.image_model] (R-72 plugin agnostik serbest string, default nano-banana, override edilebilir); project-config[brand_identity.image_style] profile-aware switch (Principle 2 enum 5-value: clean-illustration / product-photo / diagram-screenshot / location-photo / esnek); project-config[brand_identity.default_hero_url] DURUR #5 fallback URL. R-71 (8K ultra realistic prompt enrichment) + R-72 (nano-banana default model) + R-73 (1200x675 hero size + manual upload trail) + R-74 (multi-skill collaborative upload-instructions Section B) + R-75 (`<picture>` LCP optimization + loading="eager" + fetchpriority="high") + R-76 (AVIF + WebP + JPG format cascade) sentinel enforce. also_use: | new-blog skill collaborative output (Section B append upload-instructions.md, R-74 paterni); existing blog hero image regenerate (cheap AI artifact / low-res anime drift detect, R-71 enforcement); brand_identity güncelleme sonrası hero image refresh (image_style profile boyutu değişmişse). do_not_use: | Yeni blog için (new-blog kullan, image plan tetikler); FAQ enhance için (faq-optimization kullan); content prune / redirect / delete için (content-remediation kullan); image upload için (R-74 manual upload, otomatik publish YASAK, auto-upload WordPress credential exposure + plugin agnostik Ticimax/Ideasoft farklı API).
version	1.0
status	wip
category	production
inputs	{"project_slug":{"type":"string","required":true,"description":"Workspace proje slug (resolves projects/{slug}/master.xlsx + outputs/images/ + outputs/blog/{slug}/)."},"new_content_plan_id":{"type":"string","required":true,"description":"master.xlsx[new_content_plan] row id reference (kolon A); image_prompt + alt_text + content_type source (Phase 10 +3 col)."},"image_kind":{"type":"string","required":false,"default":"hero","description":"Image kind — currently hero only (enum: [hero], R-73 1200x675 fixed). Future: inline_section (Wave 3 deferred). Schema-first override: schema-frontmatter inputs[].properties whitelist 4 field additionalProperties=false; enum description'a taşındı (W-F3 D1 paterni reuse)."}}
outputs	["outputs/images/{slug}-{kind}.webp","outputs/images/{slug}-{kind}.avif","outputs/images/{slug}-{kind}.jpg","outputs/blog/{slug}/upload-instructions.md","_state/events.jsonl"]
consumes	["init-project:projects/{slug}/master.xlsx#new_content_plan","init-project:projects/{slug}/project.config.json#content_settings.image_model","init-project:projects/{slug}/project.config.json#brand_identity.image_style","init-project:projects/{slug}/project.config.json#brand_identity.default_hero_url","rules:rules/content-quality.md","rules:rules/content-html-discipline.md","rules:rules/content-seo-discipline.md","rules:rules/content-eeat-discipline.md","rules:rules/content-llm-discipline.md","rules:rules/content-update-discipline.md","templates:templates/content/upload-instructions.template.md"]
produces	["new-blog"]
triggers	{"manual":["/pseo-generate-images"],"natural_language":"\"hero image üret ve 8K ultra realistic prompt enrichment uygula\",\n\"blog görseli oluştur ve nano-banana image_model default kullan\",\n\"nano-banana image gen yap ve 1200x675 hero size R-73 zorla\",\n\"8K ultra realistic image üret ve cheap AI artifact preempt et\",\n\"hero image regenerate et ve image_style profile-aware switch uygula\"\n","hooks":[],"scheduled":[]}
mcp_tools	{"required":["mcp__higgsfield__generate_image"],"optional":["mcp__higgsfield__job_status"]}
budget	{"uses_paid_mcp":true,"estimated_credits":5}
autonomy	{"confidence":"HIGH","requires_approval":true,"safe_auto_execute":false}

generate-images — production skill (Phase 11 Wave 2 W-F5)

Hero image üretim skill'i. master.xlsx[new_content_plan].image_prompt

• alt_text source; project-config[content_settings.image_model] consume (R-72 plugin agnostik); project-config[brand_identity.image_style] profile-aware switch (Principle 2 enum 5-value); R-71/R-72/R-73/R-74/ R-75/R-76 sentinel enforce. Higgsfield MCP runtime call (Claude tool registry); fallback default_hero_url (DURUR #5). Pillow Image format cascade (WebP + AVIF + JPG, R-76); <picture> tag LCP optimization (R-75); upload-instructions.md Section B append (R-74 multi-skill collaborative output).

The skill is READ-ONLY with respect to master.xlsx (new_content_plan allowed_writers is null — F-1 schema authority). It never calls transaction.append, transaction.update, or transaction.delete against the workbook. The only state mutation is the audit append to _state/events.jsonl (event_kind=work per ADR-020 + events.schema event_kind discriminator; event_type=manual per events.schema enum).

🔴 KRİTİK — Plugin Agnostik MCP Boundary (F-16, Süleyman Seçenek D)

Worker (this skill) does NOT write to .mcp.json. Higgsfield MCP is not a plugin dependency — it is configured at VS Code user-level (Süleyman responsibility). When the plugin is given to another user, Higgsfield availability is NOT mandatory; the skill gracefully falls back via DURUR #5.

Boundary contract:

Concern	Location	Discipline
.mcp.json	Project-local (4 servers: gsc, dataforseo, ScraplingServer, sf)	Plugin agnostik; Higgsfield ABSENT (F-16 intact)
Higgsfield MCP	VS Code user-level (Süleyman setup)	Per-user, not per-plugin
HIGGSFIELD_API_KEY	Süleyman .env (12-factor app)	Never committed; never in plugin
mcp__higgsfield__generate_image	Claude tool registry runtime call	Worker calls if available; DURUR #5 if absent

STOP rule: if a worker proposal requires writing the Higgsfield server entry into .mcp.json, the proposal violates F-16 plugin agnostik boundary and MUST stop. R-72 (image_model serbest string + override edilebilir) is the architectural twin of this boundary at the schema level.

External dependency preflight (P1-13 registry contract). Higgsfield is declared in mcp-tool-registry.json under external_user_dependencies.higgsfield (NOT under servers) — the governance representation of a user-level external MCP that must already exist in the operator's own client environment. Step 4 preflights mcp__higgsfield__generate_image in the Claude tool registry; if the external tool is absent the skill does NOT crash — it routes to DURUR #5 (graceful default_hero_url fallback + a clear "Higgsfield MCP not configured at user-level" message). This is what lets the registry advertise a required-but-not-plugin-installed tool without tripping the .mcp.json↔registry server-key invariant (F-24).

Foundational Principles Enforcement (3-Layer)

The 3 üst-prensip (Phase 10 rules/content-quality.md#foundational-principles) gate this skill end-to-end. No alt-rule (R-01..R-148) overrides them.

Principle 1 — Truth-Verifiable Content (R-27, Süleyman 5x vurgu)

Statement (image context). alt_text + image caption verifiable; uydurma claim YASAK. Image içerik claim'leri ile actual generated image arasında consistency required (alt_text "doktor muayenehanesi" diyorsa generated image medical clinic resmi olmalı, generic stock photo değil).

3-katman defense:

• Layer 1: skill prompt explicit "uydurma alt_text yasak" — Step 7 upload-instructions.md Section B render öncesi alt_text source verify (manuel kayıt new_content_plan.alt_text, AI invent yasak).
• Layer 2: alt_text source verify — master.xlsx[new_content_plan] row.alt_text consume (manuel typed by Süleyman / new-blog collaborator). AI alt_text invention forbidden.
• Layer 3: image_prompt R-71 8K ultra realistic enrichment — low-res anime drift detect; cheap AI artifact preempt.

Failure mode. RED (image discard, alt_text inconsistency) — DURUR #6 trigger.

Principle 2 — Profile-Aware Enforcement

Statement. Skill behavior project.config.json[profile] enum'una göre değişir; image_style boyutu profile-driven (project-config schema image_style enum, 5-value).

Enum 5-value switch (Step 2 + Step 3):

profile value	image_style	Aesthetic intent
ymyl	clean-illustration	Medical/legal authoritative aesthetic; sterile composition
e-commerce	product-photo	Transactional intent; product-centric studio lighting
b2b-saas	diagram-screenshot	Technical clarity; UI/diagram screenshot composition
local-service	location-photo	Local SEO authority; on-location authentic photography
portfolio	esnek	Flexible — creative discretion

Failure mode. RED — image_style enum dışı bir değer (DURUR #3); profile resolve etmezse worker AMBER warning + esnek fallback.

Principle 3 — AI Suistimal Önlemi (Anti-Cheap-Content)

Statement (image context). AI'ın doğal cheap content padding davranışını image generation'da preempt et.

3-pattern gate (Step 3 + Step 4 implements):

• R-71 8K ultra realistic prompt enrichment — low-res / anime drift preempt (Süleyman explicit "8K ultra realistic" tone preference).
• R-72 image_model nano-banana default (cheap AI artifact preempt; plugin agnostik override allowed via content_settings.image_model serbest string).
• Image content cheap padding yasak — decorative-only image RED; fact-conveying / entity-illustrating image expected.

Failure mode. AMBER warning (auto-correct prompt enrichment attempt) → RED fail (manuel revise) on 2x AMBER same pass.

Routing — 7-Step Workflow

Step names are stable identifiers across runs.

Step 1 — read_master_xlsx_row

master.xlsx[new_content_plan] row read where id == {input.new_content_plan_id}. Workbook opens with openpyxl.load_workbook(read_only=True) (F-1 allowed_writers=null).

new_content_plan 14 columns consumed (Phase 10 additive bump): id, title, url_slug, primary_keyword, monthly_volume, assigned_cluster, target_word_count, priority, created_date, tivl_tag, lifecycle_status, image_prompt, alt_text, content_type. Step 1 surface fields: image_prompt + alt_text + content_type

• url_slug (output filename slot).

DURUR #1. image_prompt empty (R-71 prompt eksik) → REJECT; init-project / new-blog skill paterni new_content_plan row populate etmeli.

Step 2 — read_project_config (Profile-Aware Switch + R-72 image_model)

projects/{slug}/project.config.json parse:

• content_settings.image_model — R-72 plugin agnostik serbest string (default nano-banana, override edilebilir). Schema: {type: string, default: "nano-banana", description: "R-72. Plugin agnostik (Higgsfield veya equivalent), per-project override."}
• brand_identity.image_style — project-config schema 5-enum profile-aware (clean-illustration / product-photo / diagram-screenshot / location-photo / esnek).
• brand_identity.default_hero_url — DURUR #5 fallback URL (R-77 image alt_text fallback).
• profile (singular, schema v1.2 W-F1) — Principle 2 switch driver.

Profile-aware image_style resolve:

if profile == "ymyl":           image_style := "clean-illustration"
elif profile == "e-commerce":   image_style := "product-photo"
elif profile == "b2b-saas":     image_style := "diagram-screenshot"
elif profile == "local-service":image_style := "location-photo"
elif profile == "portfolio":    image_style := "esnek"

When brand_identity.image_style is set explicitly, the explicit value wins (override-friendly); profile-mapped value is the default when the field is null.

DURUR #3. image_style enum 5-value dışı (project-config brand_identity.image_style enum violation) → REJECT.

Step 3 — r71_prompt_enrichment (8K Ultra Realistic)

R-71 enrichment pipeline (Süleyman explicit "8K ultra realistic" tone preference; cheap AI artifact / low-res anime drift preempt):

enriched_prompt :=
    "{image_prompt}, "
    "8K ultra realistic, sharp focus, professional photography, "
    "{image_style}, "
    "1200x675 landscape, hero image composition, "
    "high detail, vibrant colors, "
    "no anime/cartoon style, no low-res artifacts"

image_style is the profile-aware suffix from Step 2 (5-enum switch). Tone keywords (8K ultra realistic, sharp focus, photographic) are R-71 mandatory; AI suistimal preempt directives (no anime/cartoon style, no low-res artifacts) are P3 mandatory.

Step 4 — 🔴 mcp_higgsfield_runtime_call (Plugin Agnostik MCP Boundary)

Worker DOES NOT write to .mcp.json. Worker calls mcp__higgsfield__generate_image from the Claude tool registry at runtime (F-16 Süleyman Seçenek D + R-72 plugin agnostik twin).

mcp__higgsfield__generate_image(
    model        = content_settings.image_model,   # default "nano-banana"
    prompt       = enriched_prompt,                # Step 3 output
    size         = "1200x675",                     # R-73 hero only
    aspect_ratio = "landscape",
)

DURUR #4. Higgsfield response size != 1200x675 (R-73 hero size violation) → REJECT.

DURUR #5 — Higgsfield MCP unavailable. Claude tool registry'de mcp__higgsfield__generate_image YOK (VS Code user-level setup eksik veya HIGGSFIELD_API_KEY env yok):

• Fallback: brand_identity.default_hero_url consume (project-config'den).
• AMBER warning emit: image generation skip.
• upload-instructions.md Section B "manual upload required, fallback hero used" mark.
• events.jsonl event_type=manual (events.schema enum), event_kind=work (ADR-020 event_kind discriminator).

DURUR #2. Higgsfield budget aşımı (estimated_credits 5 cap) → scripts/budget/check_budget.py blocked.

Plugin agnostik discipline. Worker:

• DOES NOT write mcpServers.higgsfield into .mcp.json (F-16 violation; STOP rule).
• DOES NOT hardcode HIGGSFIELD_API_KEY (Süleyman .env, 12-factor).
• DOES use mcp_tools.required: ["mcp__higgsfield__generate_image"] in frontmatter (advisory — worker reports the dependency; user-level MCP availability is not enforced via .mcp.json).

Step 5 — pillow_format_cascade (R-76)

Generated raw image (PNG/JPEG byte stream) Pillow library ile convert:

Format	Output path	Rationale
WebP	outputs/images/{slug}-{kind}.webp	Modern, geniş browser support (~30% smaller)
AVIF	outputs/images/{slug}-{kind}.avif	Next-gen (~50% smaller); Safari 16+
JPG	outputs/images/{slug}-{kind}.jpg	Legacy fallback (universal)

R-76 enforcement: 3 format hepsi yazılmalı; tek format yasak.

from PIL import Image
img = Image.open(raw_bytes_io).convert("RGB")
img.save(f"outputs/images/{slug}-{kind}.webp", format="WEBP", quality=85)
img.save(f"outputs/images/{slug}-{kind}.avif", format="AVIF", quality=70)
img.save(f"outputs/images/{slug}-{kind}.jpg",  format="JPEG", quality=88)

Step 5b — r123_iptc_disclosure_write (R-123 AI Image Disclosure)

Format cascade tamamlandıktan sonra (Step 5), WebP + JPG outputs IPTC DigitalSourceType=TrainedAlgorithmicMedia metadata ile patch edilir (Google Merchant Center AI-image disclosure policy + 2026-05-15 AI Optimization Guide reaffirmation). AVIF silent skip — piexif format desteklemez (R-123 failure_mode AMBER, R-76 cascade'in WebP + JPG çıktıları metadata garantisi sağlar).

from scripts.util.iptc_metadata import write_ai_image_disclosure

# AVIF skip — piexif unsupported (documented R-123 failure_mode)
for output_path in [webp_path, jpg_path]:
    write_ai_image_disclosure(output_path)

R-123 enforcement (rules/content-html-discipline.md). Utility: scripts/util/iptc_metadata.write_ai_image_disclosure (5/5 unit test PASS — round-trip + idempotent + missing-file + WebP + constant-format).

Step 6 — r75_picture_tag_lcp (LCP Optimization)

R-75 <picture> tag generate (LCP-optimized hero):

<picture>
  <source type="image/avif" srcset="outputs/images/{slug}-hero.avif">
  <source type="image/webp" srcset="outputs/images/{slug}-hero.webp">
  <img src="outputs/images/{slug}-hero.jpg"
       alt="{alt_text}"
       width="1200" height="675"
       loading="eager"
       fetchpriority="high">
</picture>

LCP optimization directives:

• loading="eager" — above-fold hero (lazy load yasak).
• fetchpriority="high" — Chromium 102+ priority hint.
• Explicit width="1200" height="675" — CLS preempt (R-62 alignment).
• alt="{alt_text}" — master.xlsx[new_content_plan].alt_text consume (Principle 1 truth-verifiable; uydurma alt_text yasak).

The <picture> skeleton is rendered into upload-instructions.md Section B as the canonical reference; the skill itself does not inject HTML into outputs/blog/{slug}/article.html (new-blog skill owns that artefact, R-74 multi-skill collaborative paterni).

Step 7 — append_section_b_and_emit_provenance (R-74)

templates/content/upload-instructions.template.md Section B render

• append (multi-skill collaborative output; new-blog Section A intact; revise-content Section C intact when applicable):

## Section B — Hero Image Upload (R-74)

**Generated:** {ISO 8601 UTC timestamp}
**Image Style:** {image_style}        # project-config image_style enum
**Image Model:** {image_model}        # R-72 default nano-banana
**Files:**
- WebP: outputs/images/{slug}-hero.webp ({size_kb} KB)
- AVIF: outputs/images/{slug}-hero.avif ({size_kb} KB)
- JPG:  outputs/images/{slug}-hero.jpg  ({size_kb} KB)

**WordPress Manual Upload Steps (R-74):**
1. WordPress admin → Media Library → Add New
2. Upload all 3 format files (WebP + AVIF + JPG); WordPress
   AVIF support v6.5+ (Squoosh fallback if needed).
3. Alt text: "{alt_text}"  # truth-verifiable, P1 Layer 2
4. Insert into blog post hero placeholder
   ({{HERO_IMAGE_URL_REPLACE}} token)
5. Verify <picture> tag rendering (R-75 LCP optimization)
6. PageSpeed Insights LCP < 2.5s + CLS < 0.1 measure
   (R-65 + R-62 alignment)

_state/events.jsonl append (audit, events.schema + ADR-020 compliance):

Field	Value
schema_version	"1.0" (events.schema const)
event_kind	"work" (ADR-020 event_kind discriminator)
event_type	"manual" (success AND DURUR #5 fallback) — events.schema 12-enum
event_id	UUID v4
timestamp	UTC ISO 8601
project_id	{input.project_slug} (events.schema pattern ^[a-z][a-z0-9-]*$)
note	image_generated: kind={image_kind} model={image_model} plan_id={new_content_plan_id} (note-encoded provenance)
actor	agent:generate-images (canonical code block)
image_kind	{input.image_kind} (default hero)
image_model	{content_settings.image_model} (R-72 plugin agnostik)
image_style	{brand_identity.image_style} (project-config image_style enum, profile-aware)

Canonical events_writer invocation (schema-first override 16'ıncı uygulama — event_type=manual + note=image_generated paterni; content_new schema'da url + url_normalized + after + pillar mandatory koşar, standalone hero regenerate scope'unda yok → faq-optimization + content-remediation precedent paterni reuse):

import os
import sys
import hashlib

sys.path.insert(0, os.getcwd())

from scripts.state import events_writer

# Synthetic task_id mint (whats-next paterni reuse, schema regex ^T-[0-9]{4,}$)
hex_token = hashlib.sha1(str(new_content_plan_id).encode()).hexdigest()[:4]
synthetic = 90000 + (int(hex_token, 16) % 9999)
task_id = f"T-{synthetic}"

# WORK event — schema-first override (manual + note=image_generated)
# DURUR #5 (Higgsfield unavailable) durumunda da aynı paterni: event_type=manual.
events_writer.append_work(
    project_id=project_slug,
    event_type="manual",
    task_id=task_id,
    note=(
        f"image_generated: kind={image_kind} model={image_model} "
        f"plan_id={new_content_plan_id}"
    ),
    actor="agent:generate-images",
    primary_source="manual",
)

The synthetic task_id mint paterni preserves provenance traceability via the note field (plan_id={new_content_plan_id}) without coupling the skill to a master_task row that may not exist (new_content_plan can be 1:N to master_task; init-project may delay master_task population). Schema-first override 16'ıncı uygulama: rules/events-writer.md Section 4 matrix generate-images → content_new + image_generated direkt match gibi görünse de events.schema URL-bearing work event constraints (url + url_normalized + after + pillar REQUIRED) standalone hero regenerate scope'unda satisfy edilemez; manual + note=image_generated schema authority + intent preservation = doğru fit.

DURUR Conditions (6)

#	Trigger	Resolution
1	master.xlsx[new_content_plan].image_prompt empty (R-71 prompt eksik)	REJECT — new-blog / init-project row populate etmeli
2	Higgsfield budget aşımı (estimated_credits 5 cap)	scripts/budget/check_budget.py blocked
3	image_style project-config 5-enum dışı	REJECT — schema authority cascade fix verify
4	Higgsfield response size != 1200x675 (R-73 hero size violation)	REJECT — model param mismatch, retry with explicit size param
5	Higgsfield MCP unavailable (Claude tool registry'de YOK; F-16 plugin agnostik)	Fallback brand_identity.default_hero_url + AMBER + event_type=manual
6	P1 truth-verifiable: alt_text uydurma claim detect	RED — image discard, alt_text source verify

Plugin-Agnostic Discipline

• 🔴 .mcp.json Higgsfield server entry YASAK (F-16 Süleyman Seçenek D; R-72 architectural twin). Worker STOP rule on this proposal.
• No project slug hardcoded inside this skill (demo-dental, demo-furniture, demo-hvac, demo-petcare, demo-shop, demo-tires, demo-construction, demo-agency are FORBIDDEN tokens — pytest grep verifies).
• image_model resolves at runtime from content_settings.image_model (R-72 serbest string, override edilebilir; default nano-banana).
• image_style resolves at runtime from brand_identity.image_style (project-config image_style 5-enum) or profile-mapped default.
• default_hero_url resolves at runtime from brand_identity.default_hero_url (R-77 fallback hero, DURUR #5).
• Per-call / per-url credit shape names (Phase 7 lesson, ADR-028 anti-pattern) MUST NOT appear — budget.estimated_credits (number) is the only allowed shape.

WCAG 2.1 AA Accessibility

• <img alt="..."> zorunlu (R-77; non-decorative hero image; alt_text source master.xlsx[new_content_plan].alt_text, uydurma yasak).
• Explicit width + height attributes (CLS preempt, R-62).
• h3 visible focus state — <picture> çevresi <figure> wrap edilecekse (new-blog skill owns), figcaption h3 hierarchy preserve.
• focus-visible CSS hint (R-39 axe-core simulate): <picture> interactive element değil; ancak çevresel <a> link wrap edilirse :focus-visible outline mandatory.

Error Handling

Failure	Mode	Action
new_content_plan.image_prompt boş	RED	DURUR #1 — REJECT, prompt source eksik
Higgsfield budget exhausted	RED	DURUR #2 — check_budget.py blocked
image_style enum dışı	RED	DURUR #3 — schema cascade fix verify (v1.2 W-F1)
Higgsfield response 1200x675 size mismatch	RED	DURUR #4 — REJECT, retry with explicit size
Higgsfield MCP runtime registry'de YOK	AMBER	DURUR #5 — default_hero_url fallback, event_type=manual
P1 truth-verifiable alt_text uydurma	RED	DURUR #6 — image discard, alt_text source verify
Pillow AVIF encode fail (libavif missing)	AMBER	WebP + JPG only fallback; AVIF skip mark in Section B
<picture> LCP measurement > 2.5s post-publish	AMBER	manual measure post-upload, R-75 enforcement check

References

• rules/content-quality.md (Foundational Principles 1+2+3, R-01..R-148).
• rules/content-html-discipline.md (R-71 8K ultra realistic, R-72 nano-banana default, R-73 manual upload, R-74 multi-skill collaborative output, R-75 LCP optimization, R-76 format fallback, R-77 alt text, R-123 IPTC DigitalSourceType disclosure).
• rules/content-seo-discipline.md (R-65 LCP < 2.5s ranking signal, R-62 CLS preempt).
• rules/content-eeat-discipline.md (R-44 source verify; image authority signal).
• schemas/project-config.schema.json v1.2 (content_settings.image_model serbest string default nano-banana; brand_identity.image_style 5-enum; brand_identity.default_hero_url format=uri).
• schemas/master-excel.schema.json (new_content_plan 14 col Phase 10 +3 col additive: image_prompt, alt_text, content_type).
• schemas/events.schema.json (event_kind=work ADR-020 discriminator; event_type 12-enum events.schema: manual for success AND DURUR #5 fallback).
• schemas/skill-frontmatter.schema.json (8 required fields, category enum 8 values, status enum 3 values, mcp_tools.required advisory).
• .mcp.json (4 servers: gsc, dataforseo, ScraplingServer, sf; Higgsfield ABSENT — F-16 plugin agnostik boundary intact).
• ADR-020 (event_kind=workflow vs work routing).
• ADR-028 (per-call / per-url credit shape anti-pattern).

Wave-Out Note

This skill is status: wip until Phase 11 Wave 2 acceptance gate; active bump deferred to the post-Wave 2 closeout. Manager owns the status flip + governance audit. Atomic commit Wave 2 6'ıncı kanıt hedef (W-F3 + W-F4 + W-F5 paralel atomic dispatch + drift sıfır).