feat: add Ollama embedding provider + unify provider selection (AGE-1)

[saldestechnology]

Implements AGE-1. Adds Ollama as a third embedding backend (alongside fastembed/local and OpenAI), so users get high-quality embeddings that run fully offline and free — and makes real end-to-end ctx smart relevance testing viable (the fastembed model can't run deterministically in CI; a local Ollama can).

What's new

• OllamaProvider (src/embeddings/ollama.rs) → Ollama /api/embed.
  • Host OLLAMA_HOST (default http://localhost:11434, bare host:port normalized), model OLLAMA_EMBED_MODEL (default nomic-embed-text), optional OLLAMA_API_KEY bearer for remote hosts.
  • reqwest client + timeouts + retry/backoff mirroring openai.rs; actionable errors when the daemon is down or the model isn't pulled (ollama pull …).
• Dynamic dimension — Ollama's dimension is model-dependent (nomic-embed-text 768, qwen3-embedding:8b 4096, …), so it's probed from the model on construction (sync from_env, plus from_env_async for the MCP runtime), not a compile-time constant.
• Unified provider selection — replaced the use_openai: bool duplicated across ~6 call sites with a Provider enum + one build_provider factory and a --provider <local|openai|ollama> flag on embed, semantic, smart, similar, plus a provider field on the MCP smart_context tool. --openai kept as a deprecated alias.
• Mismatch detection — a shared warn_index_mismatch warns when the query provider/dimension differs from the index (different providers/models occupy different vector spaces → re-embed).

Incidental fix

ctx embed's --force claimed -f, colliding with the global -f/--format. This only trips clap's debug-build assertion, so no prior release invocation hit it — but the new debug test surfaced it. Dropped the short (other subcommands' --force are already long-only; the watch commands use --force).

Tests + docs

• Unit tests: host/model resolution + /api/embed response parsing (success, dimension mismatch, model-not-found, empty).
• tests/ollama_smart_e2e.rs — a real ctx smart --provider ollama relevance test, gated behind CTX_TEST_OLLAMA=1 + a daemon reachability probe, so it's a no-op in normal CI (no charges, no downloads). Asserts relative ranking (not vectors) so it survives model drift. Verified locally against a live daemon with qwen3-embedding:8b: ctx smart "parse solidity contract source into a syntax tree" selects src/parser.rs above the off-topic files.
• README + docs site (configuration, code-intelligence) document the providers and env vars.

Verification

All gates green: cargo fmt --check, cargo clippy --all-targets -D warnings, full suite (298 lib incl. new Ollama unit tests + all integration suites). The gated e2e passes live and skips cleanly when CTX_TEST_OLLAMA is unset.

Acceptance criteria (AGE-1)

• OllamaProvider implements EmbeddingProvider with configurable host + model
• Dimension determined dynamically from the model
• embed / similar / smart / MCP analysis can all use --provider ollama
• Provider selection unified into a single factory (no scattered use_openai)
• Provider/dimension mismatch against an existing index is detected & surfaced
• Unit tests for request/response parsing + gated integration test behind a running Ollama
• Docs updated (embeddings/semantic-search section)

Run the e2e locally: CTX_TEST_OLLAMA=1 cargo test --test ollama_smart_e2e -- --nocapture (set CTX_TEST_OLLAMA_MODEL to override the default qwen3-embedding:8b).

[saldestechnology]

Added project config support: an optional committed .ctx/config.toml ([embedding] provider/model/host) so a repo can set a default embedding backend without flags. Precedence is CLI flag > env > config > default; .gitignore keeps config.toml tracked while ignoring the rest of .ctx/. Verified end-to-end: with provider = "ollama" in the file, ctx semantic/smart use ollama with no --provider flag.