Cherry-picks all unique features from FarelRA, hnewcity, BcGee,
Yorick-Ryu, severity1, gaoshan5211, Echoxiawan, and leikaiwei forks.

Changes:
- Account system: Circuit Breaker, sticky behavior, lazy init, multi-account
- API region: auto-detection from profile ARN with env-var override
- Auth: support for social login (Google/GitHub) and Enterprise Kiro IDE
- Config: ACCOUNTS_CONFIG_FILE, MODEL_CACHE_TTL, PROXY_AUTH_DISABLED
- Converters: responses API format, document/PDF extraction, thinking blocks
- Debug logger: per-request archival, mode=all/mode=errors
- Extensions: tool name aliasing system
- HTTP: SOCKS5 proxy support, connection timeout detection
- Models: dynamic resolution, fallback models, hidden model support
- Parsers: inverted model names support, Cursor flat format
- Profile ARN: centralized profile_arn_for_payload() helper
- Routes: /v1/responses endpoint, /v1/usage endpoint, PROXY_AUTH_DISABLED
- Streaming: first-token timeout with retry for both APIs
- Token counting with tiktoken
- Truncation recovery system
- Network error classification with user-friendly messages
- PDF text extraction from document content blocks (both APIs)
- Live model catalog refresh on /v1/models endpoint
- Content-Type: migrated from x-amz-json-1.0 to application/json

Fixes:
- Session fixture: set OS env vars so config module reloads preserve
  temp credential paths (fixes 79 test errors in full-suite runs)
- All 1673 unit tests passing

…talog

Replace the hardcoded native-reasoning tables with per-model schemas fetched
from the Kiro control plane. ListAvailableModels on management.{region}.kiro.dev
advertises each model's additionalModelRequestFieldsSchema, which declares whether
the model supports native adaptive thinking and exactly which effort levels it
accepts. Binding to that ground truth fixes two problems in the static approach:
requesting native effort on models that don't support it, and sending effort
values (e.g. "xhigh") outside the server's strict enum.

- kiro/native_reasoning.py: new registry that parses each model's advertised
  additionalModelRequestFieldsSchema into an effort descriptor (schema type,
  valid effort enum, default) and clamps a requested effort onto the advertised
  set (xhigh -> max when the model only offers low/medium/high/max).
- account_manager: fetch the catalog from the control plane (management host,
  ListAvailableModels JSON-1.0 RPC) for all accounts, including runtime-endpoint
  ones, and register the advertised schemas on init and TTL refresh. Falls back
  to static models on failure. Removes the runtime-endpoint fetch short-circuit.
- converters_core.build_native_effort_fields: consult the catalog registry first
  (static config table only as fallback) and clamp effort to the advertised enum.
- auth: add management_host; config: point KIRO_MANAGEMENT_HOST_TEMPLATE at
  management.{region}.kiro.dev; http_client: allow per-request extra_headers so
  the catalog RPC can set its x-amz-target.

Verified live against a Kiro subscription: management host returns HTTP 200 for
origin=AI_EDITOR, and only claude-opus-4.6/sonnet-4.6 advertise effort
(low/medium/high/max), which the registry picks up automatically.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018gszhX9bogSALZ4S9AR1Vq

feat(reasoning): drive native reasoning effort from the live model catalog