feat(tools): hermes-borrowed quality wins — error/summary caps + sharper tool descriptions

[HongmingWang-Rabbit]

Summary

Three small wins from the hermes-agent design survey, bundled because each is too small for its own PR but they all improve the priority adapters (claude-code + hermes) immediately.

1. Cap telemetry fields inside the helper, not at call sites

Hermes' `error_detail` cap (4096 chars) lives once, inside the helper. We had it at the call site of `tool_delegate_task` only — a future caller adding error_detail elsewhere could paste a giant stack trace and DoS `activity_logs`. Move the cap into `report_activity` so every caller benefits.

• `error_detail`: 4096 chars (hermes' value — multi-frame stack trace fits, 100 errors/5min still <500KB)
• `summary`: 256 chars (one-liner ceiling for canvas card)
• `response_text`: NOT capped — that's the agent's user-visible reply; truncating would silently drop content. Pinned by a negative-pin test.

2. Sharper MCP tool descriptions

Hermes' `delegate_task` says "WAIT for the response"; `delegate_task_async` says "Returns immediately." LLMs pick the right variant from the description — ambiguity costs accuracy.

• `commit_memory` now states it APPENDS (each call creates a row), and that GLOBAL requires tier-0
• `recall_memory` now states it's case-insensitive substring search with no pagination, returns ALL matches, and that empty query is cheap

3. Bigger wins filed as tasks

• Task security(auth): PATCH /workspaces/:id unauthenticated — allows name/role/tier/parent_id/runtime mutation #120: per-workspace tool enable/disable menu in Canvas Config (M effort, the big user-flow win)
• Task security(C19): GET /workspaces topology protected in running binary but not in repo code — divergence risk #121: model-string passthrough — let adapter resolve providers (depends on chore(template): add 4 evolution crons — ecosystem / plugins / template / channels #87)
• Task fix(security+scheduler): IPv6 SSRF gap + scheduler unit tests [supersedes #111, #112] #119: hermes architectural patterns (event-log, declarative skill compat, observability config block)

Tests (4 new)

• `error_detail_capped_at_max` — 50K input → 4096 stored (load-bearing assertion)
• `error_detail_under_cap_passes_through` — short value not padded/truncated
• `summary_capped_at_max` — 1000 chars → 256 stored
• `response_text_NOT_capped` — negative pin: 20K reply stays 20K so capping discipline doesn't accidentally extend to user-visible content

Verification

• 1312/1312 workspace pytest pass (was 1308, +4 new — exact match)
• No code path changed for response_text or any payload field other than error_detail/summary

🤖 Generated with Claude Code