Load the kernel whole and decide delegate routes at dispatch (5.0.1)

[mzored]

Summary

Release 5.0.1.

Observed defect.

• Claude Code middle-truncates Bash output above about 10,000 characters. The 5.0.0 kernel was 25,024 characters, and agents usually load it with cat.
• In 11 field sessions the middle of the kernel was cut away: routing, the delegate brief contract, and custody. A re-read is cut the same way.
• tracked-work.md (12,076 characters) was cut once.
• Separately, dispatches left the delegate's model unset and inherited the lead's route, and delegates' own nested delegates inherited it unseen.

Change.

• Kernel. Rewritten to 9,614 characters (52 lines). It keeps the invariants: trust, request grants, iteration, protected actions, owner decisions, custody, and completion. Method moved to the reference whose trigger fires at the act, or was deleted as a frontier-model default. The disposition table is in docs/decisions.md.
• Dispatch. Every delegate's model and effort is set through host controls, and an unset control is a routing decision held to the same test. The brief states whether the delegate may dispatch delegates, and at what route and bound.
• references/delegation.md. Adds the nested fan-out rule. references/tracked-work.md. Compressed to fit one view.
• Gate. scripts/check.py fails any runtime Markdown file under plugins/skiphow/ whose numbered view (chars + 7 × lines) exceeds 10,000. The reopen condition is recorded in docs/decisions.md.
• Carriers. Version carriers, eval anchors, CHANGELOG, docs/evidence.md, and docs/prior-art.md are updated.

No owner interface, authority boundary, default side effect, public skill name, activation scope, or package layout changes.

Review

• Round 1. Claude subagents in two lanes:
  • Opus 5, invariant preservation against v5.0.0: 11 findings.
  • Sonnet 5, carriers and anchors: 3 findings.
• Round 2. 7 findings. Round 3. 2 findings.
• Every finding was confirmed and fixed; none were refused. Codex review was not run, per owner instruction.

Validation

Check	Result
scripts/check.py (391 tests)	PASS
git diff --check	PASS
Codex schema validation	PASS
Claude schema validation	PASS
Clean installs, invocation, activation, continuity	UNVERIFIED
Behavioral contract suite	UNVERIFIED (headless Claude auth expired; see docs/evidence.md)

🤖 Generated with Claude Code