fix(skills): key skill-command cache per platform to prevent cross-platform leaks

[draix]

Problem

Closes #14536

agent/skill_commands.py maintains a process-global _skill_commands dict. The first platform to call scan_skill_commands() seeds this cache with its own platform-specific disabled-skill view. Subsequent platforms (e.g. Discord after Telegram) hit get_skill_commands(), find the cache non-empty, and return the wrong skill set — inheriting the first platform's disabled list.

Root Cause

# Before
_skill_commands: Dict[str, Dict[str, Any]] = {}  # single global cache

def get_skill_commands():
    if not _skill_commands:   # any non-empty cache → returned as-is for ALL platforms
        scan_skill_commands()
    return _skill_commands

_get_disabled_skill_names() correctly resolves per-platform disabled lists via HERMES_SESSION_PLATFORM, but the cached result was shared across all platforms.

Fix

	Before	After
Cache structure	Dict[str, Any] (flat)	Dict[str | None, Dict[str, Any]] (per-platform)
scan_skill_commands()	no platform arg	accepts platform=None, resolves and uses as cache key
get_skill_commands()	no platform arg	accepts platform=None, checks per-platform slot
Platform resolution	always from env	explicit arg → HERMES_SESSION_PLATFORM → HERMES_PLATFORM → None

All existing call-sites use the no-arg form and continue to work — they now automatically resolve platform from session context.

Tests Added (TestPlatformAwareCache)

• test_different_platforms_get_independent_caches — Telegram and Discord see their own disabled lists
• test_platform_none_uses_global_disabled_list — global disabled list respected when platform is None
• test_rescan_updates_platform_cache_entry — rescan replaces the per-platform slot

[alt-glitch]

Competing fix with #14570 for #14536 — this PR keys cache per-platform rather than invalidating on change. Maintainer should pick one.

[teknium1]

Thanks for working on this. Closing in favor of #18739 which salvages #14570.

Applying this PR to main and running the issue's repro returns an empty result for both platforms — scan_skill_commands() now calls _get_disabled_skill_names(resolved_platform), but that function in tools/skills_tool.py takes no argument. The resulting TypeError gets swallowed by the outer except Exception in scan_skill_commands(), so scanning silently returns empty:

telegram: []
discord: []

The architectural approach (per-platform keyed cache, passing platform through explicitly) is cleaner than the rescan strategy in #18739, but getting there would require updating _get_disabled_skill_names() to accept a platform argument in tools/skills_tool.py AND agent/skill_utils.py. Happy to revisit if you want to open a follow-up with that full change — for now shipping the less-invasive rescan fix.