evals: the three 5.8.0 documents

[Masterplanner25]

Gate 10 (creator validation), Stage 5 (post-publish) and Stage 6 (downstream sweep) for 5.8.0, plus the consumer republish Stage 6 turned up.

Stage 5 found a real defect — filed as #691

A workflow step body that calls a function in an imported module stops executing at that call, and the run is reported as successful.

Four different symptoms from one construct, depending on the shape of the imported module — silent truncation, Stack underflow, Cannot call non-function: nil, and a spurious nodus-retry is required error. That spread is the signature of stack/dispatch corruption rather than a logic error.

The sharpest evidence: a module defining only no_loop works, a module defining only in_loop works, and a module defining both fails when no_loop is called.

The silent case is the severe one:

STEP RAN
failed: []
steps: {}

print("got \(v)") never ran, no error was raised, and the run reported no failures. The control — same file, let v = 7i instead of the module call — prints got 7.

Not a regression

Identical on the published 5.7.1 and 5.8.0 wheels, tested in clean venvs from outside the repo. 5.8.0 is where it starts to matter, because retry.until is a module function whose documented home is a step body.

Why nothing caught it

tests/test_retry_until.py, all eleven new Gate 10b probes, and the guide example all run their code inside fn main(). The full suite, nine gate phases and 83 probes were green on a feature that does not work in the position its own documentation points at.

The generalisation worth keeping: a construct documented for use inside a step body has to be tested inside a step body. The step-body path and the top-level path are two paths, and a test on one says nothing about the other — the recurring shape, again.

Gate 10 found two things in the probe work itself

1. One of my own new probes could not fail. cli() takes argv including the program name, so cli(["workflow", "--help"]) consumed "workflow" as argv[0] and asserted against the general usage — it would have passed against a CLI with no such verb.
2. The 5.4.0 stale-prose pattern fired on a true sentence — a no longer executes the file row in CLAUDE.md's table. Its self-check could not have caught this: the pattern was written out twice, and the self-check validated a copy.

Stage 6

• All 6 published companions still admit 5.8.0; all 11 match what they published.
• nodus-run-action was stale — its README pin is what new users copy. Republished as v1.0.9, floating v1 moved and verified against the remote (rev-parse on an annotated tag returns the tag object, not the commit). Flag cleared with fingerprint and published updated in the same commit.
• nodus-vscode in step — 5.8.0 adds no keywords.
• All 13 sibling checkouts clean.

Note on the last commit

The eval-record claim in ECOSYSTEM_READINESS_ASSESSMENT.md declares points_at, so it only becomes checkable once docs/evals/v5.8.0/ exists — it passed before the eval documents landed and failed immediately after. That is the check working as designed: the version string and the file it names are verified together.

🤖 Generated with Claude Code

https://claude.ai/code/session_013uyCMZqPCwvXnjkgc1mSFJ