refactor(skills): decompose netclaw-operations + reconcile identity-vs-memory routing

[Aaronontheweb]

Summary

Decomposes the 1,148-line netclaw-operations skill into a hybrid progressive-disclosure layout (a variant of #1115) and reconciles the identity-vs-memory routing rule to openspec/specs/netclaw-agent-memory.

Part 1 — Decompose netclaw-operations (#1115)

SKILL.md drops 1,148 → 346 lines. Safety-critical / high-frequency guidance stays inline (tool-argument validation, large-output handling, the full approval-prompts section, identity routing); the long-tail manuals move into references/*.md, surfaced on demand via skill_read_resource.

Why hybrid, not the full split #1115 sketched: I gated the decomposition on whether the local-tier model (Qwen3.6-35B-A3B) reliably does the skill_load → skill_read_resource two-hop. It does ~80% of runs (3/5 clean, 1/5 after thrashing, 1/5 bypassed to a web fetch of the source). For long-tail detail that's fine; for distilled guidance with no external fallback a missed hop = lost guidance — so the high-value content stays inline.

Mechanism verified end-to-end (no runtime/feed changes needed):

• Skill auto-activation keys off the frontmatter description (unchanged) — the split does not affect activation (SkillRegistry.GenerateIndex/Search).
• references/*.md are auto-discovered (SkillScanner.EnumerateResources) and surfaced by skill_load; skill_read_resource reads them (scoped to references//scripts//assets/).

Part 2 — Reconcile identity-vs-memory routing

Per openspec/specs/netclaw-agent-memory, durable user facts & preferences are memory documents (store_memory); identity files (SOUL.md/AGENTS.md) define the agent. The binary AGENTS.md memory-triage table, skill-authoring, and netclaw-operations had all routed user "preferences" → SOUL.md, contradicting the spec and netclaw-memory. The model correctly used store_memory and failed the old eval. Aligned all four sources.

Part 3 — Evals

• Added skill_progressive_disclosure — verifies the skill_read_resource second hop via a reference-only detail (the reminder auto-disable threshold/alert name).
• Repurposed memory_identity_preference_routing to assert the spec-correct invariant: routed to memory, never an identity-file edit. 5/5 on the local-tier model.

Version bumps: netclaw-operations 2.17.0, netclaw-memory 1.5.0, skill-authoring 1.7.1. (2.17.0 rather than 2.16.0 because #1481 independently took 2.16.0 on dev; this PR is rebased past it and folds its sub-agent-liveness wording into references/scheduling.md.)

Verification

Full rebuild + 55-case eval suite vs Qwen3.6-35B-A3B (spark1): tool-calling 8/8, identity 4/4, grounding 3/3, multi-turn 7/7 — no regressions from the split or the always-loaded AGENTS.md change. memory_identity_preference_routing 5/5 (refined assertion). Remaining suite movement is small-sample model nondeterminism in untouched categories. skill_progressive_disclosure lands ~70–80% — the intended ongoing signal for the two-hop read (and the reason for the hybrid layout).

CI

This is the first multi-file skill, so I hardened publish_skills.yml: the post-upload validation previously only spot-checked the first skill's SKILL.md URL. It now also fetches a files[] (reference) URL from the generated manifest and asserts HTTP 200 — self-verifying that progressive-disclosure reference files actually publish to R2 end-to-end, not just the top-level SKILL.md.

Verified the full publish chain on this branch: generate-skill-manifest.sh emits netclaw-operations v2.17.0 with all 10 references/*.md in files[] (each with path/sha256/sizeBytes/url) and an upload-list entry per file; publish_skills.yml uploads every upload-list entry to R2; the consumer (SkillFeedManifest / SystemSkillSyncService) reads files[] with matching field names and downloads + sha256-verifies each reference atomically.

Closes #1115.