fix(desktop): hash spawn config as the env receives it, not raw record fields

[wesbillman]

Follow-up to #1602 (restart-required badge), parked behind the spawn_hash persona-resolution rewrite that landed in #1618.

Defect

The spawn-config hash digested raw record values while the spawn env receives normalized ones, so three field edits badged running agents even when a restart would run the exact same process:

Field	Spawn env behavior	Hash behavior (before)
respond_to_allowlist	BUZZ_ACP_RESPOND_TO_ALLOWLIST set only in allowlist mode, from validate_respond_to_allowlist output (trim/lowercase/dedup)	raw list, hashed in every mode
max_turn_duration_seconds	BUZZ_ACP_MAX_TURN_DURATION written with default filled in	raw Option — None ≠ Some(default)
mcp_toolsets	BUZZ_TOOLSETS falls back to the default set	raw Option — None ≠ explicit default

Equal spawned env + unequal raw field = spurious badge.

Fix

Hash what the env receives:

• Allowlist: hashed only when respond_to == Allowlist, normalized via the same validate_respond_to_allowlist spawn uses. On validation error the raw list is hashed — spawn rejects invalid lists, so a stamped hash always came from a valid one and the invalid edit correctly compares unequal (badge = honest "restart would change/fail" signal).
• Max turn duration: hashed with DEFAULT_AGENT_MAX_TURN_DURATION_SECONDS filled in, matching the env write.
• Toolsets: hashed through a new DEFAULT_MCP_TOOLSETS const in types.rs, which also replaces the string literal in runtime.rs's spawn path — hash and env now share one source of truth and cannot drift.

respond_to mode itself stays hashed unconditionally: toggling modes is spawn-visible and must badge.

Tests

+8 in spawn_hash/tests.rs (17 total): equivalence AND sensitivity in both directions for all three fields — dormant-list edits under owner-only don't badge, real allowlist content edits do; normalization-equivalent edits don't, content changes do; None ↔ explicit default doesn't, non-default values do.

Full desktop-tauri suite: 1048 passed / 0 failed. Clippy --all-targets clean, fmt clean, file-size check clean.

Cross-reviewed and approved by Brain in #centralize-personas-and-agents (verified against build_respond_to_env and the runtime.rs env writes; independently ran the suite).