* Record per-skill turns/cost and surface in PR body + summary

Adds two capture steps that parse claude-code-action's execution
log (`/home/runner/work/_temp/claude-execution-output.json`) right
after each skill invocation, BEFORE the next one overwrites the
shared file. Exposes `turns`, `cost_usd`, and `permission_denials`
as step outputs for downstream use.

Surfaces the data in two places:

1. PR body: new "Run cost" subsection inside the upstream-release-
   docs marker block. Per-session rows plus a Total row when both
   sessions reported. Applies to both pull_request and
   workflow_dispatch runs.

2. workflow_dispatch summary comment: adds Turns and Cost columns
   to the existing step table, plus a Total row summing both
   sessions.

Useful for tracking per-release spend ($6 baseline) and catching
regressions -- e.g. a release that suddenly takes 10x the turns is
visible at a glance rather than requiring a drill-down into the
Actions Step Summary.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Reorganise PR body and comments for scannability

Ground-up redesign of everything the workflow posts on a PR.
Previous output read as three authors writing on the same page
(log formatter + release notes + systems docs), with no visual
hierarchy and a lot of process lore competing with actionable
info. Feedback after PRs #779 and #780.

Key structural shift: the PR body is the single source of truth
for state. Comments are point-in-time events only. Process lore
goes behind <details>.

Changes:

**PR body** (rewritten end-to-end)
- New H2: "Docs update for <project> <tag>" replaces the
  process-descriptive "Content additions by upstream-release-docs".
- New "At a glance" table with 6 rows: Upstream (linked compare),
  Hand-written changes, Reference assets, Gaps, Release
  contributors, Action required. Answers the 5 questions a
  reviewer actually asks in < 5 seconds.
- Single top-of-section alert (at most one), ordered by severity:
  CAUTION for autogen-path drift, NOTE for NO_CHANGES.md, NOTE
  for silent run. Never stacked.
- Autogen drift promoted from inline [!WARNING] paragraph to
  [!CAUTION] at the top with a bulleted revert list -- it's a
  merge-blocker, should look like one.
- "Review guidance" paragraph deleted. The one operational claim
  ("reference files are auto-synced") is in the At-a-glance
  "Reference assets" row.
- "Release contributors" section dropped when everyone got
  auto-assigned (GitHub's sidebar already shows them). Only
  rendered as "Additional release contributors" when overflow
  @-mentions need a target.
- "Run cost" table kept visible, not buried in <details>, per
  request.
- Process narrative ("Two Claude Opus sessions...", auto-synced
  paths, Helper-prompt convention) collapsed into a single
  <details> block at the bottom.
- Compare-failure warning moved from a top-of-section [!WARNING]
  block into the At-a-glance "Release contributors" cell.

**Bootstrap PR body**
- Collapsed from two paragraphs to one line ("Manually dispatched
  by @x. Workflow run: ..."). Mechanism narration deleted.

**Pre-skill placeholder comment** (workflow_dispatch only)
- From a two-paragraph explanation + parenthetical meta-note
  ("this comment exists because track_progress isn't supported
  on workflow_dispatch") to a single line: "Generating docs for
  <project> <tag>... (run)".

**Post-skill summary comment** (workflow_dispatch only)
- Table replaced with a one-line status: "Done · 93 turns · $6.04
  · run · see PR body for details". On failure, identifies which
  step died. Duplicative detail (conclusions, commit counts) was
  already in the PR body.
- Dropped "Skill commits produced" row -- the at-a-glance
  "Hand-written changes" cell covers it.

**Augmentation failure comment**
- Wrapped in [!CAUTION] alert, compressed to two lines with the
  exact retry gh command inline (not a hand-wavy "retry via the
  workflow" pointer).

**Autogen-touch step (`detect-autogen`)**
- Now emits `touched` output (comma-separated paths) directly
  instead of pre-formatting a warning block. The Augment step
  composes the alert itself. Cleaner separation of concerns.

Unchanged: reviewer assignment logic, skill prompts, autofix
step, concurrency config.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>