Multi-file reports, one-carrier token math, and GitHub-enrichment probing

[bgmcmullen]

Summary

Docs/skills only; no runtime code paths change. Three threads across the plugin
skills, mirrored in the claude/ and codex/ workspaces.

1. Multi-file report output

The four report skills (adoption, improvement, security, spend) now emit a
progressively-disclosed one-pager (hypaware-reports/<date>-<name>.md) plus
one markdown file per detail section, instead of a single file. The two
standing quality rules now bind the section files, not just the summary:

• Section files are analysis, not inventory - each argues one claim and opens
  with its takeaway; a section that is only a stat table is folded back into the
  one-pager rather than shipped as a page.
• No scope apologies (in any file) - never print "descriptive only" /
  "no recommendations here" / routing disclaimers.

2. Token math reconciled to the one-carrier rule (LLP 0035 / LLP 0026)

Usage rides exactly one row per response (the last assistant part; non-carrier
parts are null), so a plain SUM over assistant attributes.usage rows is
correct with no dedup. Verified empirically against local recordings (19,639
usage-bearing rows): usage never appears on more than one part per message, and
the old dedup query returns a byte-identical total to a plain SUM
(8,732,420 in / 15,109,690 out both ways).

• Removed the stale "usage repeats across parts / a plain SUM overcounts ~3x"
  guidance from the spend and improvement reports and the hypaware-query skill.
• Replaced the spend SQL with a plain SUM.
• Switched the defensive-dedup key from conversation_id to session_id
  (conversation_id is null for Claude and only separates threads within a Codex
  session).

3. GitHub-enriched graph + mandatory reach probing

• hypaware-graph: documents server-side GitHub enrichment (Actor / Issue /
  PullRequest / Review nodes, enriched Repo / Commit / File, new edges) and the
  bridge-ready natural keys that converge a session's HEAD commit and a PR's
  referenced commit on one node.
• adoption report: GitHub reach is no longer a guessed "if enriched"
  condition - step 1 must probe for a github.t0 projector every run, and
  when present, step 2 must compute each gateway's real reach (repos/PRs its
  work landed in, whether it drew review), attributing sessions via cwd. When
  absent, the report states "checked - no GitHub enrichment present"; the
  phrasings "not assessed" / "not queried" for reach are banned.

Housekeeping

• De-em-dashed the four report skills and the adoption edits per repo code style.
• Reworded the adoption opener to drop the "the one descriptive report / no
  actions" framing.

Refs: LLP 0035 (token-usage-normalization), LLP 0026 (claude-native-granularity),
LLP 0028 (context-graph-enrichment), LLP 0032 (github-llm-graph-bridge).

🤖 Generated with Claude Code