feat(0.10.1): G1-G4 + C1 + H1-H4 follow-up sweep

[tbitcs]

0.10.1 follow-up sweep — diversity / capabilities / phase routing / trace seal / token threading + docs

Closes the remaining G1-G4, C1, and H1-H4 todos from the 0.10 multi-agent sprint plan.

What landed

G1 — Diversity guard on agents add

ProfileStore.diversity_warnings walks the profile population and warns
when the reviewer (or architect) shares a provider family with the
coder. New PROVIDER_FAMILIES table groups OpenAI-family endpoints
(openai-compat, azure-openai) and Ollama-family backends
(llamacpp, vllm, lmstudio) so a self-hosted vLLM coder + Ollama
reviewer is correctly flagged as same-family. The CLI prints yellow
warnings (non-fatal) and the --json output now includes
diversity_warnings.

G2 — Capability filter

ProfileStore.filter_by_capability(capability) plus a new
specsmith agents list --capability code-review --json flag. The
extension counterpart (AgentsClient.filterAgentsByCapability) ships
in the matching specsmith-vscode PR.

G3 — phase next auto-routes

Advancing the AEE phase now pins a synthetic phase:active route to
the new phase's preferred profile (and seeds the canonical
phase:<key> entry on first advance). The runner can flip the whole
session by listening for one activity instead of teaching the user
seven agents route set commands.

G4 — TraceVault seal on /agent

In-chat /agent <id> writes a decision seal chained into
.specsmith/trace.jsonl so every per-turn profile pin is auditable.
Best-effort: a read-only filesystem / missing project root must never
break the chat loop.

C1 — Token threading

Each provider driver now returns (text, _UsageDelta) and surfaces
real token counts:

• Ollama — prompt_eval_count + eval_count from the final done message
• Anthropic — final_message.usage.input_tokens / output_tokens
• OpenAI — stream_options.include_usage makes the final SSE chunk carry the usage block
• Gemini — usage_metadata.prompt_token_count / candidates_token_count

When the SDK omits usage, a 4-chars/token heuristic fills in so the
TokenMeter chip is never zero. Counts flow through
ChatRunResult.tokens_in/out/cost_usd into AgentState.credit() and
the per-profile by_profile bucket.

H1-H4 — Docs

• docs/site/agents.md — preset → route → per-session → BYOE walkthrough
• docs/site/quickstart.md — reproduction script + GIF placeholder
• docs/site/vscode-extension.md — eight new commands documented (sister-PR scope)
• README — multi-agent + BYOE elevator pitch up top

Verification

• ruff check src/ tests/ — All checks passed!
• ruff format --check src/ tests/ — 148 files already formatted
• pytest -q — 448 passed, 1 skipped
• python -m specsmith.cli api-surface | diff - tests/fixtures/api_surface.json — clean (no surface drift)

Out of scope / follow-up

• A5 release tags — landed once this PR + the matching specsmith-vscode#? PR merge
• Marketplace publish — separate manual step gated on VSCE_PAT