Add postmortem skill

[Multipixelone]

Summary

Adds the postmortem marketplace skill — write a blameless incident postmortem
from the exports an incident leaves behind (chat channel export, pager timeline,
fix PRs, metric data). Every timeline entry cites its source; gaps are marked
instead of filled; trigger, root cause, and contributing factors are kept
separate (an unestablished root cause stays unknown); handles are replaced by
roles so no individual is named; and action items carry an owning role and a
calendar date. A local linter flags blameful phrasing, speculation, leaked
names, uncited timeline rows, and vague action items before publication.

What's included

• skills/postmortem/SKILL.md and three references (timeline reconstruction,
  blameless writing, contributing factors).
• skills/postmortem/scripts/timeline.py — collect / render / check,
  Python standard library only.
• skills/postmortem/pyproject.toml — declares zero runtime dependencies.
• skills/postmortem/assets/postmortem-template.md and a complete worked
  example under assets/example/ (five raw exports through the finished doc).
• marketplace.json catalog entry and a README Current Status bullet.

Dependencies & safety

• tier: python_uv, recipes: [] (nothing to install), system_capabilities: [python, uv].
• has_helper_scripts: true, runs_install_script: false,
  network_at_runtime: false — runs entirely on local files, no network.

Verification

• Byte-for-byte reproducible example. collect regenerates
  assets/example/timeline.json and render regenerates
  assets/example/postmortem.md identically to the committed files;
  check --strict on the finished example reports 0 findings.
• All repo checks pass: scripts/validate_skills.py, check-jsonschema
  against schemas/marketplace.schema.json, markdownlint-cli2, typos,
  and gitleaks.
• The uv run --project … python scripts/timeline.py form shown in SKILL.md
  was smoke-tested against the pyproject.
• collect/check were exercised against malformed inputs (non-object
  records, ambiguous compact timestamps, quoted multi-line CSV fields, and a
  gap whose anchoring event is excluded) to confirm the tool reports problems
  rather than crashing or silently dropping data.