feat(skills): experiment, audience-review, ad-fatigue-check, and incident-postmortem workflow skills

[hyoshi]

feat(skills): experiment, audience-review, ad-fatigue-check, and incident-postmortem workflow skills

Four new bundled workflow skills, each with a packaged copy under
mureo/_data/skills/ and a byte-identical repo-root mirror under skills/:

• experiment (A/Bテスト設計・実行・評価): turns an ad-hoc change into a
  designed experiment — a falsifiable hypothesis (variable / expected
  effect / success metric), exactly one variable, a Goal-bound metric,
  and a sample/duration floor from _mureo-learning. Setup on Meta via the
  native split-test tools (meta_ads_split_tests_create/_get/_list/_end;
  cells are pre-existing ad sets); Google Ads has NO experiment tool, so
  it offers the RSA-asset-A/B (google_ads_rsa_assets_analyze) and
  duplicated-campaign (google_ads_campaigns_create +
  google_ads_ad_performance_compare) paths with their caveats spelled out
  and explicit operator acknowledgment required. Records per-variant
  baselines in action_log (observation_due = designed end date), forbids
  peeking-based decisions before the window closes (sequential-testing
  bias), evaluates each variant with mureo_outcome_evaluate to a
  winner / no-difference / inconclusive verdict (inconclusive is valid),
  and offers to save the finding via /learn. Persists report="experiment".

• audience-review (オーディエンス・配置レビュー): audits current targeting
  against the STRATEGY.md Persona — Google Ads demographic/audience/device
  reads (google_ads_demographic_targeting_list /
  google_ads_audience_targeting_list / google_ads_device_analyze), Meta
  audiences, ad-set targeting, and placement/audience analysis
  (meta_ads_audiences_list / meta_ads_analysis_placements /
  meta_ads_analysis_audience / meta_ads_insights_breakdown) — flags
  Persona mismatches (contradicting spend, missing segments, placement
  waste), and executes only the mutations mureo supports
  (google_ads_bid_adjustments_update, meta_ads_audiences_create_lookalike,
  meta_ads_ad_sets_update), presenting demographic/audience criterion and
  Meta placement exclusions as honest manual UI steps. action_log
  observation_due 14 days. Persists report="audience".

• ad-fatigue-check (クリエイティブ疲弊チェック): scores active ads
  FATIGUED / WATCH / FRESH. Meta frequency is DERIVED (no frequency field
  is exposed) as impressions ÷ reach from meta_ads_insights_report; CTR
  trend from LAST_7_DAYS vs (LAST_14_DAYS − LAST_7_DAYS) subtraction; CPM
  drift; Google display/RSA CTR trend from
  google_ads_ad_performance_report and RSA asset-level fatigue via
  google_ads_rsa_assets_audit/_analyze. Rubric: frequency >= 3.5 + CTR
  decline >= 20% w/w = FATIGUED; one signal = WATCH; learning insights
  override. Noise guards: >= 1000 impressions per window; single-day dips
  never count. Routes fatigued ads to /creative-generate or
  /creative-refresh; pausing (meta_ads_ads_pause /
  google_ads_ads_update_status) is an approval gate. Persists
  report="fatigue".

• incident-postmortem (インシデント振り返り): closes the learning loop
  after a /rescue — makes NO ad-platform writes. Scans action_log
  (mureo_state_get) for TURNAROUND_RESCUE / command="/rescue" entries,
  reconstructs the timeline, evaluates closed-window actions with
  mureo_outcome_evaluate, runs platform/site/measurement/external
  root-cause analysis (5-whys), writes a structured postmortem document
  (./postmortems/YYYY-MM-DD-.md on Code, inline on Desktop),
  distills 1-3 insights via /learn, and — with approval — proposes
  preventive STRATEGY.md ## Guardrails (mureo_strategy_set). Logs the
  postmortem to action_log. Persists no report.

Honest-limitation calls (no invented tools): Google Ads has no
campaign-experiment tool; Meta insights expose no frequency field
(derived from impressions/reach); demographic/audience criterion and Meta
placement exclusions have no mureo mutation tool (manual UI steps).

Registration/tests/docs: EXPECTED_PACKAGED_SKILLS 21->25; _BUNDLE_NAMES
+4 alphabetical; extended tests/test_new_workflow_skills.py to cover the
4 new skills (mirror byte-parity, frontmatter version from pyproject,
Before-you-start pointer, >=3 verified tool-name references, approval-gate
language, report_set keys for the three that persist, and a
no-ad-platform-writes assertion for incident-postmortem); README /
README.ja workflow tables +4 rows (20 operational skills); getting-started
operational count 16->20 and upload list +4 in both languages; CHANGELOG
Unreleased Added entry.

Claude-Session: https://claude.ai/code/session_01AbvhGbHZoK8VqP3kR9nWWp