fix(gateway): multiplexed profiles use their own prefill_messages_file and provider_routing

[AhmetArif0]

What does this PR do?

On a multiplexed gateway, a routed profile's turns run with the launch (default) profile's prefill_messages_file and provider_routing instead of their own.

GatewayRunner._init_runtime_settings reads both once at boot, outside any profile scope, into self._prefill_messages / self._provider_routing, and every agent the gateway builds takes that copy:

• gateway/run_turn_runner.py::_build_fresh_agent → prefill_messages=runner._prefill_messages
• gateway/run_turn_runner.py::run_sync → pr = runner._provider_routing (→ providers_allowed/ignored/order, provider_sort, provider_require_parameters, provider_data_collection)
• gateway/run_turn.py::_run_background_task_inner (/background) → pr = self._provider_routing

So a secondary profile's own only / ignore / order never apply, and a provider_routing.data_collection: deny set on that profile is silently dropped: its requests go out with the default profile's data-collection preference.

This contradicts multi-profile-gateways.md ("Per-turn runtime settings follow the routed profile … never from the profile the gateway was launched under"). It is also a gap in #108453: that PR listed _load_prefill_messages among the per-turn reads it moved to the active home (819517f), but the loader is only ever called at boot, so the change never took effect for a routed turn.

The fix reads both where the turn runs, inside the routed profile's _profile_runtime_scope. That matches how the same turn already resolves reasoning and service tier, and how the ephemeral system prompt was fixed for the same boot-snapshot problem in #89161. Cron already reads both per job, and the TUI/Desktop backend per agent build; the gateway was the only surface that snapshotted them.

Related Issue

No issue filed. Found while checking the per-turn settings #108453 moved to the routed profile.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• gateway/run_turn_runner.py: _build_fresh_agent calls runner._load_prefill_messages(); the turn computes pr = runner._load_provider_routing() next to the per-turn reasoning / service-tier resolution.
• gateway/run_turn.py: /background computes pr = self._load_provider_routing() (it already runs inside _profile_scope_for_source).
• gateway/run.py: _init_runtime_settings no longer snapshots _prefill_messages / _provider_routing. No production code reads them any more, and keeping a launch-profile copy around would invite the same bug.
• gateway/run_config_loaders.py: _load_prefill_messages docstring now says it resolves from the active gateway home and must be called per agent build (it said ~/.hermes/).
• website/docs/user-guide/multi-profile-gateways.md: adds provider_routing (including data_collection) and prefill_messages_file to the list of per-turn settings that follow the routed profile.
• tests/gateway/test_routed_profile_prefill_routing.py (new).

How to Test

Two profiles, both prefill_messages_file: prefill.json, with different provider_routing (the routed one adds data_collection: deny). The test drives the real GatewayRunner._run_agent → TurnRunner.run_sync → _build_fresh_agent path, and the real _run_background_task, inside the routed profile's _profile_runtime_scope (the scope _profile_scope_for_source enters under multiplexing). The runner also carries the launch profile's boot copy, so the test proves a turn no longer uses it.

Routed profile's agent kwargs	main	this branch
prefill_messages	default-PREFILL ❌	beta-PREFILL ✅
providers_allowed	["default-only"] ❌	["beta-only"] ✅
provider_data_collection	None (dropped) ❌	"deny" ✅
/background providers_allowed / data_collection	["default-only"] / None ❌	["beta-only"] / "deny" ✅
Launch profile's own turn	default values	default values (unchanged)

Mutation check. Each change was reverted on its own with the test file kept:

• prefill read → the main-turn test fails on prefill_messages
• turn pr → the main-turn test fails on providers_allowed
• /background pr → the background test fails

With all three restored, both tests pass.

pytest tests/gateway/test_routed_profile_prefill_routing.py -q

Notes / scope

• Standalone gateway: the values are now read when an agent is built, not only at boot. So an edit to provider_routing / prefill_messages_file reaches newly built agents without a restart, the same as service_tier and fallback_providers today. A cached agent keeps what it was built with. Rebuilding cached agents when routing changes is what fix(gateway): reload provider routing and bust cached agents on routing changes #32088 proposes; this PR does not touch the cache signature.
• Torn writes: a per-turn read does not drop routing while config.yaml is being written. load_user_config_effective serves the last good parse for a broken file (in-process, then the .good backup).
• Env override unchanged: HERMES_PREFILL_MESSAGES_FILE stays a process-wide override, like HERMES_EPHEMERAL_SYSTEM_PROMPT.
• Heads-up for feat(gateway): show configured OpenRouter routing after model switch #110814 (open, P3): it reads getattr(self, "_provider_routing", {}). After this change that attribute no longer exists, so the display would show {}. Calling self._load_provider_routing() fixes that, and also shows the routed profile's routing under multiplexing.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate. fix(gateway): reload provider routing and bust cached agents on routing changes #32088 reloads routing per turn for cache-busting on the pre-split run.py; it does not cover prefill or multiplexing.
• My PR contains only changes related to this fix
• I've added tests for my changes
• I've run the relevant tests.
  • Full tests/gateway + related files at base 9bb2ea1b5c: 16 failures on this branch, all of which also fail on main at the same base. main has one extra flaky failure (test_control_socket::test_collect_fleet_versions_falls_back_to_state_file). No new failures.
  • After rebasing onto current main: every test file that builds a runner with these attributes (38 files, plus the new one and test_personality_routed_profile.py) passes, 258 tests.
  • ruff check is clean on the changed files, and so is scripts/check-windows-footguns.py.
• I've tested on my platform: macOS 26.5 (Apple Silicon), Python 3.14

Documentation & Housekeeping

• I've updated relevant documentation (multi-profile-gateways.md, loader docstring)
• cli-config.yaml.example: N/A (no config keys added or changed)
• CONTRIBUTING.md / AGENTS.md: N/A
• Cross-platform impact considered: no platform-specific code. The test writes its files with an explicit encoding="utf-8".