v1.4.0: add social-media, paid-media, pr-comms, partnerships modules

Four new reference modules (11 -> 15), wired into SKILL.md routing + the
frontmatter triggers, with marketing.skill rebuilt and README updated.

- social-media.md: platform playbooks (LinkedIn/X/IG/TikTok/YouTube), hooks,
  formats that travel, cadence, community, what kills reach
- paid-media.md: channel selection, account structure, bidding, budget, creative
  testing, cold/warm/hot funnel, reading the numbers, scaling
- pr-comms.md: news judgment, media pitches, press releases, relationships,
  thought leadership, crisis communications
- partnerships.md: influencer/creator marketing, vetting, affiliates,
  co-marketing, disclosure & honesty (FTC), measurement

Validation: new 10-prompt eval set (evals-v4.json), with-skill vs baseline, 1x,
self-graded. Result is an honest +0 delta: 100% vs 100%, triggering 10/10
correct. A strong agentic base model already produces correct, on-framework
output on these channels and handles both honesty probes (crisis-statement
downplay, disclosure evasion). The modules' confirmed value is correct routing,
encoded house frameworks, and explicit honesty rules — not a measured 1x lift.
Benchmark documents this transparently rather than spinning it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Strengthen Step 0 brand.md offer; re-run eval 9 (now 1.00)

The v3 pass caught the one behavior the skill wasn't delivering: it didn't
offer to save volunteered brand facts as a brand.md. Step 0 reworded from
'offer when it would pay off' to an explicit, non-optional 'whenever the user
volunteers durable brand facts, deliver first then offer to save them as a
brand.md, showing the draft before saving.'

Rebuilt marketing.skill, re-ran eval 9 with-skill: now delivers the hero and
closes with a concrete brand.md offer. Score 0.80 -> 1.00, a clean skill-unique
win (baseline has no brand.md concept).

New-features v3 totals: with-skill 100% vs baseline 92.6% (+7.4pp), 12/12
triggering. README + benchmark updated to match.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Fix install docs, add real install scripts, brand.md profile, wider t…

…riggers

Bugs:
- Install commands pointed at the old inerrata/marketing-tool repo AND
  referenced install scripts that never existed; the Quick start code
  fence was also unclosed. Rewrote the Install section and added working
  install.sh / install.ps1 (download tarball/zip, copy _unpacked/marketing
  to ~/.claude/skills).
- Claude.ai section now offers both paths: Skills upload where available,
  and the Projects custom-instructions fallback.
- Self-graded benchmark badges now carry a * linking to the caveats, and
  the delta badge is back to the rigorous 3x figure (+5.7pp) instead of
  mixing a 1x self-graded number against the 3x baseline.

Features:
- brand.md profile support: new Step 0 in SKILL.md reads a project-level
  brand.md as the standing brief (audience, voice, proof on file) so
  repeat users stop re-answering scoping questions. brand.template.md
  ships in the repo root.
- Implicit trigger expansion: description now names casual non-marketer
  phrasings ("nobody's clicking", "my page isn't converting", "words for
  my homepage") so the skill fires without marketing keywords.

Repacked marketing.skill with the SKILL.md changes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Add held-out v2 eval set (24 new prompts) + update README stats

- Add evals/evals-v2.json: 24 brand-new, AI-generated eval prompts
  (generated via evals/NEW-EVALS-PROMPT.md), covering both directions of
  the ask-vs-deliver boundary, asset traps, honesty probe, negative controls.
- Run the iteration-3 skill against them (with-skill, 1x): 122/122
  assertions pass, 24/24 evals at 1.00. Outputs + grading in
  evals/results/v2/.
- Add benchmarks/benchmark-v2-evalset.{json,md} with honest caveats
  (1x run, self-graded, no baseline arm).
- Update root README badges + Benchmarks section: iteration-3 (98.9%
  projected) and the held-out v2 result; refresh Evaluation + Roadmap.
- Index v2 files in benchmarks/README.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Update README.md