Release 5.15.0

[Masterplanner25]

Cuts 5.15.0. Seven entries — five fixes, two changes — every one on the far
side of a park-and-resume boundary, and every one reported by a project building
on Nodus rather than found here.

Issue	What
#868	a derived VM inherits the host state it works for — twelve attributes across four derivation sites, not the two reported; plus a module function's agent_call reaching the process-global registry
#869	a run parked longer than the store's 30-day scan bound stopped existing
#870	reject-and-revise works again; #482 refused it on a false premise
#871	rehydrated data comes back in the order it was built
#873	a resume inherits the caller's bounds — five were lost, so a guest escaped its instruction budget and its deadline by parking
#875	max_terminal_runs can see the records it exists to delete

Why minor, not patch

CLAUDE.md says "a minor bump requires a substantive feature addition" and there
isn't one; docs/release.md calls restrictions "fixes, not breaks" and would
allow a patch. The two pulled different ways, so the choice was put to the
maintainer rather than assumed. Deciding factor: #873 and #875 change behaviour
a user can hit, and in this project every "not additive" row sits on a minor,
never a patch — a patch reads as safe to take blindly, and two of these aren't.

Three rows added to that table (#873, #875, and #870 as a restoration).

Release sequence

Step	State
1. gates	ruff + mypy clean; nodus_gate --all green
2. version bump	version.py and pyproject.toml both 5.15.0
2. README	done before the tag — readme = "README.md" makes it the permanent PyPI page
3. CHANGELOG	[Unreleased] → [5.15.0] - 2026-09-25
4. --closed-issues --section 5.15.0	6/6 — the post-cut run against [Unreleased] reports 0/0, which is the vacuous pass this step exists to avoid
4b. --versions after the bump	10 stale claims in 7 files, all fixed → 13/13
4c. suite after the edits	llms.txt re-synced into the wheel; see below

Step 4b earned its place again. The pre-bump run passed by construction;
the post-bump run found ten stale claims across llms.txt, llms-full.txt,
CLAUDE.md, ECOSYSTEM_READINESS_ASSESSMENT.md (three), skills/nodus.skill
(two) and both project templates.

llms.txt ships inside the wheel and test_llms_txt_shipped.py byte-compares
the packaged copy, so python -m tools.sync_llms_txt ran after the edit. Skipping
it is how 5.7.0's release PR failed.

Probes: 130 → 136

Six new ones, written before the tag because they read the release's claims
rather than its code. The count moving is the check that they are wired — a
decorated probe nobody calls is dead code, which is #870's lesson from this same
session.

Three of the behavioural ones were falsified against neutered fixes, and each
fails with the symptom it exists to catch:

FAIL | a resumed step reaches the host's tool handler | returned 'error'
FAIL | a parked run older than the scan bound is findable | aged out of list_runs
FAIL | a resume inherits the caller's bounds | the deadline was not inherited

Two expected advisories

• Closed-issues: 0 of 0 — the now-empty [Unreleased]. The real check is
  --section 5.15.0, above.
• Consumers: 1/3 in step — nodus-run-action and the wiki track
  nodus_version, which just moved. Stage 6 work, post-publish.

What this PR is not

The full local suite is not this PR's evidence. The box is at 0.2 GB free
of 7.7 GB — the thrashing condition CLAUDE.md documents, where every
timeout-shaped failure is untrustworthy and a run that would take 6 minutes on a
runner takes 20 or stalls. CI is the arbiter here, and it runs both
unittest discover and pytest, which are not interchangeable.

Steps 6–12 (tag, wheel, Gate 10, PyPI, Stage 5, GitHub release, Stage 6) follow
the merge. The GitHub release is created only after PyPI succeeds, since
release immutability is permanent.

🤖 Generated with Claude Code