docs: trim CLAUDE.md, move companion detail to its own file

[Masterplanner25]

CLAUDE.md is loaded into context every session, so length there is a running cost. 1,889 → 1,619 lines, with no lesson dropped.

Companion detail → docs/ecosystem/COMPANION_REPOS.md

Eleven per-companion sections sat in three separate regions of the file, interleaved with unrelated material, and the key-file table listed all the same paths a fourth time. They are one document now, reached from a fourteen-row index.

Every gotcha survives — the two C:\codev directories and which one is not the wire repo, the nodus-mcp egg-info pitfall, the VS Code PAT organization scope, the nodus-flow rename rule, and the module-vs-distribution collision that deleted a package's classes with pip reporting success both times.

Two sections were saying the same thing twice

SemVer (146 → 85). It carried a narrative per release duplicating CHANGELOG.md, told 5.7.0's supersession once at the top and again eighty lines down, and printed the "X is current" vs "as of X" paragraph verbatim in two places.

What survives is what the changelog cannot tell you at a glance: the not-additive table — which answers "is this symptom a release, or my change?" — and the rules the release process needs. It now also records that the 5.7.0 rule is about a release a fix will supersede, and does not apply when the Stage 5 defect is pre-existing. That is the distinction 5.8.0 had to make.

Published ecosystem (116 → 94). Per-package detail that now lives in the new file.

Two version claims retired rather than re-registered

They guarded CLAUDE.md quoting version.py's and pyproject.toml's literal contents back at the reader — a third copy of a fact that already has two authorities, kept in step by a claim, while the gate's own version files agree check compares those authorities directly.

The restatement was removed instead of the claim being repointed. A claim is worth having when prose asserts something a reader acts on; it is not worth having to guard a transcription that did not need to exist. 13/13 claims pass, and version files agree still covers the real authorities.

Also updated

• BUG: a workflow step body calling an imported module's function truncates silently and the run reports success #691 is documented in the language quirks with its repro and workaround, and named in the recurring-shape section as an open candidate — deliberately not in that table, whose standard is "confirmed by reading the code" and which BUG: a workflow step body calling an imported module's function truncates silently and the run reports success #691 has not met.
• The published-ecosystem section records that re-deriving the package count only works if the list it derives from is complete, which at 5.8.0 it was not.
• The token note was wrong: the account-scoped ~/.pypirc token creates a new project on first upload, so a project-scoped token is an alternative rather than a requirement. Established publishing nodus-workflow-ai.

Facts re-measured, not carried forward

Claim	Was	Now
tests collected	2,940	3,132
gate symbols	136	138
runtime blocks	263	268
.venv vs src/	5.0.0 / 5.7.1	5.0.0 / 5.8.0

Gate

All nine phases pass — static 138/138, runtime 268/268, contracts 6/6, opcodes 28/28, consumers 2/2, shapes 0 new, versions 13/13, invariants 4/4.

🤖 Generated with Claude Code

https://claude.ai/code/session_013uyCMZqPCwvXnjkgc1mSFJ