Spectator mode: permission-gated ghost-pawn join + broadcast overlay

[Jeronyx]

Summary

• Spectator mode: a new ga_user_roles-gated -spectate <instance_id> [attack|defense] chat command that joins a live instance as a permanent no-pawn viewer — no model, no collision, no scoreboard entry, no world interaction. The optional team argument assigns PRI.r_TaskForce/Team purely so the client's own same-team HUD checks resolve; it never spawns a pawn. -unspectate reverses it and routes the player back home.
• Crash fix uncovered along the way: HandlePlayerConnected unconditionally dereferenced Controller->Pawn for post-join bookkeeping (GPawnSessions, skill reapply, spawn GAME_EVENT). Never mattered before — every prior code path eventually got a pawn. A real spectator is the first caller that keeps Pawn null permanently, which crashed the DLL the first time it was actually exercised.
• Spectator broadcast overlay: a rate-limited per-pawn health + active-effect/skill-id feed (DLL → in-memory control-server state → token-gated HTTP endpoints → a local browser-source HTML page meant for OBS). Gated end-to-end on GActiveSpectatorCount > 0 — the DLL push, the HTTP polling, and the DB reads it triggers all cost nothing when nobody is actually spectating.
  • Raw ids the DLL reports are filtered through a new OverlayIdCatalog whitelist and split into effect ids (rendered as an icon tile) vs. skill ids (forwarded but never rendered on the main card) before ever reaching the overlay page.
  • A companion skill-tree overlay page (skill-tree-overlay.html) shows all 6 players' full skill trees side by side, backed by a new SkillTreeCatalog + GET /overlay/skilltree/GET /overlay/builds endpoints. The main overlay's health card now also shows a compact "P/T1/T2" build notation (Balanced/Tree1/Tree2 point totals) per player.
  • Main overlay visual pass: restyled health bar (squared-off, bordered, gradient "depth" fill), vertical centering, a shared left-to-right class order (Assault/Medic/Robotics/Recon) applied independently to both teams so the two rows read consistently, an H-key admin-bar hide toggle for clean OBS capture, and a toggle to hide the raw HP number entirely (off by default — in-game, players only ever see the bar, never a number).
• Removed: the disabled TgPawn::IsFriendlyWithLocalPawn diagnostic hook (address was Ghidra-confirmed correct, but installing it alongside two older unverified hooks crashed a real Steam client with STATUS_ILLEGAL_INSTRUCTION, cause never root-caused) and the abandoned dinput8-flavored client diagnostic build target it needed. Left commented-out in a prior commit while under investigation; now deleted outright since it was never required by the overlay feature (which reads pre-existing pawn state server-side and needs none of this) and wasn't going anywhere.

Explicitly out of scope for this PR: an in-progress -nextplayer/-prevplayer spectator camera-cycling command was prototyped alongside this work and deliberately held back — not included in any commit here, left for a follow-up PR.

Performance note for review

The control server runs a single shared ASIO io_context across every listener — logins, inventory, beacons, chat, matchmaking, TCP sessions, and now the overlay HTTP server all share it. A slow synchronous call anywhere stalls everything else, not just that one request, so each new poll loop was sized with that in mind rather than left at whatever felt convenient during ad-hoc testing:

• GET /overlay (300ms): health + effect/skill ids, served straight out of an in-memory map (SpectatorOverlayState) populated by the DLL's push — no DB access at all on this path, which is why it can afford to poll fast.
• GET /overlay/skilltree (2s, new): the skill-tree page's live point-by-point breakdown genuinely needs a per-player DB read (GetSkillsForCharacter) every tick, so this is deliberately ~7x slower than the health poll — a couple of cheap SQLite reads every 2s for a small match, kept off the fast path on principle rather than because it's measured to be expensive.
• GET /overlay/builds (6s, new): the "P/T1/T2" summary reads through a 30s-TTL cache (SkillTreeCatalog::GetCachedBuildSummary) that's also event-driven invalidated the instant a build actually changes (profile-slot swap, skill save/respec all call InvalidateCache immediately) — so a real build change shows up within one 6s poll tick rather than waiting out the cache, while a steady-state match costs at most one DB read per player per 30s regardless of how often the page polls.
• DLL-side push is unchanged from the original review round: reuses the existing non-blocking IpcClient::Send() path (enqueue + async post, same as every other IPC message) and stays rate-limited to ~3 pushes/sec per real player pawn, filtered out immediately for bots/deployables/projectiles.
• The HTTP polling endpoints have no keep-alive (one full TCP round-trip per poll) — worth keeping an eye on under real match volume with several spectated instances at once, but the tiered intervals above (300ms/2s/6s, matched to what each endpoint actually needs to read) are the main lever already pulled to keep this affordable on a box that isn't a resource powerhouse.

Test plan

• Grant spectator role, -spectate a live instance, confirm: no pawn/model/collision/scoreboard entry, free-flight camera works
• Confirm -spectate <id> attack|defense doesn't affect matchmaking/team-balance (ga_instance_players unaffected)
• Confirm both overlay pages render correctly for a real 3v3+ match: health/effects on the main overlay, full skill trees on the skill-tree page, consistent class ordering across both teams on both pages
• Confirm the HP-number toggle defaults off and correctly swaps between showing/hiding the raw number without affecting the health bar fill itself
• Confirm a mid-match loadout-profile swap updates the "P/T1/T2" build notation within one /overlay/builds poll tick (not delayed by the 30s cache TTL)
• Load-check control-server CPU/latency on logins/inventory/beacons while both overlay pages are actively polling a live match