fix: model init, session cache, retry semantics; config accessors raise AttributeError for missing keys

[wilsonhj]

⚠️ Contains a deliberate accessor-contract change — please read first

This PR consolidates the model-initialization, session-cache, and retry-semantics fixes, and changes the config accessor contract, because the two are one atomic change set (verified empirically — see "Why config.py is in this PR" below):

• Root Config.__getattr__ missing-key behavior: KeyError → AttributeError
• Nested ConfigWrapper.__getattr__ missing-key behavior: silent None → AttributeError (breaking for downstream callers relying on the None return; item access cfg["key"] stays lenient)

The old silent-None made hasattr() always-True and getattr(x, k, default) never return its default — hiding config typos and forcing per-call-site workarounds throughout the codebase, which this PR removes.

Fixes

• Silent exception swallowing in _get_direct_model_: a bare except: pass permanently discarded all model-init errors (wrong API key, missing provider package, network failure). Now: expected fallback conditions (ImportError/ValueError) log and fall through to manual init; unexpected errors re-raise as LCELModelConfigError with the original traceback chained. Error-message construction itself can no longer raise (safe getattr for model_name).
• Provider handling: enumerated providers keep their manual init branches; non-enumerated providers are passed explicitly to init_chat_model(model_name, model_provider=...) so langchain-supported providers (mistral, groq, ollama, …) work, while genuinely unknown ones fail fast with a clear "not supported" error — and the check ordering no longer emits a misleading "falling back" warning or "API key not found" for a key that was never needed.
• Azure routing: Azure-shaped configs route to a dedicated builder before the generic fast path (which has no notion of azure_endpoint/api_version/engine); endpoint detection parses the hostname and suffix-matches Azure domains (blocks lookalike-host/path spoofing); api_version is validated with a clear error on both direct and EAS paths.
• SessionMap thread safety and correctness: is_expired() no longer raises KeyError for unregistered sessions; all reads/writes lock-guarded; singleton construction uses double-checked locking with full initialization before publication; get_valid_llm() does the existence+expiry check atomically; expires_in is captured per session (a later reconfiguration no longer retroactively changes cached sessions' expiry) and resolved before any network call.
• Retry semantics: NonRetryableError marker family — permanent config errors fail fast instead of burning max_retry attempts, strictly per-family so a broad filter can't re-enable retries for unrelated marked exceptions; transient EAS auth failures (non-200) raise the retry-eligible EASAuthError instead of a permanent error.
• Typing: wrap_llm_with_proxy accepts BaseLanguageModel (the real common ancestor of chat models) instead of the deprecated langchain.llms.BaseLLM import.

Why config.py is in this PR

The coupling is bidirectional: the rewritten consumers rely on missing-key AttributeError (root Config's old KeyError isn't swallowed by getattr(default)), and the new accessor breaks the old consumers' except KeyError paths. Splitting them produces two individually-red PRs; together the suite is fully green.

Verification

• Full unit suite on this branch: 121 passed, 0 failed — including the chains/orchestrator/MCP tests still at main's state, confirming the accessor contract was the only cross-subsystem coupling.
• Fixes main's own pre-existing test failure (test_model_with_unsupported_provider).
• pylint on touched files: 9.89/10 vs 8.24/10 baseline; mypy: no new real errors (3 new import-not-found are guarded optional provider packages).

Sequencing

Includes the shared dependency pin (needed to build standalone); that hunk rebases away once #10 merges. Best reviewed/merged after the sanitizer/orchestrator PR.

Co-authored-by: Claude noreply@anthropic.com

[wilsonhj]

Runtime verification — PASS ✅ (all 7 claims)

Verified through the public package boundary with the real Config.from_env() loader over hand-written YAMLs (CONFIG_PATH), no network calls, no test-suite rerun.

1. Accessor contract: existing nested reads work; missing top-level and nested keys both raise AttributeError with the key named; getattr(cfg, 'nope', 'fallback') and hasattr now genuinely work; ConfigWrapper item access stays lenient (None).
2. Unsupported provider fail-fast: provider: definitely-not-real → LCELModelConfigError with the cause chained; captured logging output is empty — no misleading "falling back" warning.
3. langchain-known but uninstalled provider (mistralai): clear error with the exact pip install -U langchain-mistralai instruction chained. (Cosmetic nit: the lead phrase "not supported" slightly undersells "just not installed" — the chained hint rescues it.)
4. Azure anti-spoofing: https://notopenai.azure.com.evil.example/v1 does not route to the Azure builder (proceeds as plain OpenAI); a genuine *.openai.azure.com endpoint without api_version raises the clear Azure error before any opaque pydantic failure.
5. SessionMap: is_expired() on an unregistered id returns True (no KeyError); round-trip via get_valid_llm(); singleton confirmed; staleness honored.
6. Retry fail-fast by call counters: LCELModelConfigError (NonRetryable) → exactly 1 call; plain ValueError → 3 calls (max_retry); explicitly listing a marked family re-enables retry for that family only.
7. Malformed config: missing models section → LCELModelConfigError: "No models defined in config" with the AttributeError cause chained.

Error-message quality — which is half of what this PR is about — held up under every adversarial probe.