fix: replace deprecated openai.gpt-5.5 model id; add OCIBucketBackend retry strategy

[fede-kamel]

Summary

Two production audit fixes from the 2026 best-in-class assessment.

1. Model id correction — openai.gpt-5.5 → openai.gpt-5

`openai.gpt-5.5` is not a real model in OCI's catalogue (the actual id is `openai.gpt-5`). The deprecated form had spread across 47 files in docs, examples, provider docstrings, and a hero SVG. Net effect: a new user copying the quickstart hits ModelNotFoundError on first run.

Replaced everywhere; added a regression test (`tests/unit/test_model_id_consistency.py`) that fails if `gpt-5.5` ever reappears in `src/`, `docs/`, `examples/`, or `README.md`.

2. OCIBucketBackend retry strategy

The OCI Object Storage checkpointer made every put / get / delete / list call without a retry strategy, so any transient 429 / 5xx / connection error caused an immediate save failure — a real risk in multi-region failover scenarios.

Now the backend threads `oci.retry.DEFAULT_RETRY_STRATEGY` (exponential backoff + jitter) into all six call sites by default. New `retry_strategy=` constructor kwarg lets callers override (pass `oci.retry.NoneRetryStrategy()` to disable for tests).

Six new unit tests in `tests/unit/test_oci_bucket_retry.py` mock the OCI client and assert the strategy is preserved across:

• `put_object`
• `get_object`
• `delete_object`
• `list_objects`
• `get_bucket`
• `create_bucket`

Plus the default-resolution test (`oci.retry.DEFAULT_RETRY_STRATEGY` when none passed).

Companion issues filed

The audit surfaced larger items that need their own PRs:

• Decompose Agent god class (src/locus/agent/agent.py is 2,073 LOC) #32 — Decompose `agent.py` god class (2,073 LOC)
• Redesign memory/store.py — replace NotImplementedError leakage with capability-flagged Protocol #33 — Redesign `memory/store.py` to replace `NotImplementedError` leakage with capability-flagged Protocol
• Migrate mypy ignore_errors blanket to per-line directives #34 — Migrate mypy `ignore_errors` blanket to per-line directives
• Anthropic prompt caching support #35 — Anthropic prompt caching support
• Surface OpenAI structured outputs at provider layer #36 — Surface OpenAI structured outputs at provider layer
• Ship Dockerfile + Helm chart for AgentServer deployments #37 — Ship Dockerfile + Helm chart for AgentServer deployments

Test plan

• All 3,212 unit tests pass (`hatch run test`)
• `pre-commit run --all-files` clean (ruff lint + format, mypy, codespell, markdownlint, gitleaks)
• `hatch -e docs run mkdocs build --strict` clean (zero warnings)
• DCO sign-off
• `grep -rE "gpt-5\.5" src/ docs/ examples/ README.md` returns zero matches
• New `test_model_id_consistency.py` catches any future regression