fix(agent-runtime): remediate onboarding boundary-audit findings

[Parad0x-Labs]

Context

KAS's NULLA boundary audit (Parad0x-Labs/nulla-video#2, 34 findings) reproduced onboarding-chat failures on a build that already carried the prior live-memory remediation, so the residual causes were elsewhere — the OpenClaw persona/session layer, the local-model context window, and a set of in-repo routing/grounding/guard bugs. This PR addresses the in-repo-fixable subset.

What changed

• Routing. The web0 builder fast-path no longer fires on prose ("make a five-step web0 guide"); the live-price fast-path drops the x402 stage word quote, matches markers on word boundaries, and skips a Web0/x402 topical turn, so x402/.null prompts stop returning crypto quotes; the hive create intent no longer fires on "summarize without adding new topics".
• Input. Stop appending a stale Context subject: reference onto a turn that declares its own subject ("keep this label: X") — that had corrupted the turn and the session summary into an invented lease record.
• Memory. A content-avoidance directive ("do not mention X") is no longer stored as a durable global constraint that contaminates unrelated fresh sessions, on both the rule-based and LLM-extraction paths. An explicit "from now on ..." directive still persists (intended).
• Response guards. Strip a leaked OpenClaw NO_REPLY control token (reveal the real answer behind a prefix; drop it from persisted history so it can't re-feed). Add a wallet-safety guard that replaces a response soliciting a private key / seed phrase, or claiming a payment completed, with a documented-safe correction.
• Grounding. Add docs/NULLA_SELF_KNOWLEDGE.md (injected as must-keep context) with the tight safety/capability facts; add deterministic answers for /stopx402, the default cloud-off policy, and "settlement cannot be self-claimed"; the x402 definition now names the request → quote → verify → settle → receipt flow.
• Config. Write an OpenClaw compaction reserve floor so a long session compacts instead of hard-erroring at the 32k context precheck.

Validation

• Two adversarial review passes over the diff found and fixed regressions in the new guards and regexes before merge. The wallet guard is a best-effort defense-in-depth backstop (not a complete filter) and was validated against 50 solicitation / false-payment / honest-content cases; the primary F6 defense remains the grounded self-knowledge doc.
• New regression suite tests/test_onboarding_boundary_remediation.py pins each finding, including the legit-preserved cases. Targeted suites and ruff are green locally; the full matrix runs sharded in CI.

Deferred to live validation (KAS)

• The context-window overflow's deeper levers — the tools.profile: full schema size and the local Ollama num_ctx (4096 default / adaptive-context flag) — are VRAM/runtime-sensitive and need a hardware run, not a blind config flip.
• The OpenClaw session-projection issue (multiple user text blocks merged into one) is upstream of nulla-local.
• Please re-run the boundary batch against this branch's main before merge (same loop as fix(installer): validate port health on restart and reset OpenClaw session state (KAS-049) #92).