MC redesign 6/6: calibration, stop-hazard consumers, MCP is_pitting

[VforVitorio]

Part of #550. Closes the epic, #434 and #439.

What

• B_SC re-cast as the gap-cost reduction inside the projection (measured 5.75 s, CI [3.14, 8.25], n=124); POS_GAP_S demoted to legacy-path-only; the VSC bonus stays documented as an unmeasured placeholder.
• The stop-hazard table wired as OVERCUT eligibility on the surfaces whose rivals lack is_pitting (/recommend, MCP), per the capability matrix from the gate-1 audit.
• Every doc page describing the MC updated in this PR (docs are part of the fix): docs site multi-agent + agents-api + arcade pages, the agents README, the submodule design spec. Signature tables regenerated by introspection, never by hand.
• A/B simulation report in the PR body: Lusail 2025 SC laps + a green race + the sweep argmax distribution, before vs after.

Closes #434
Closes #439

[VforVitorio]

Re-scoped rather than closed. Two of the three parts landed; the middle one was overtaken by a better fix.

Done (on main since #576):

• The MC docs, in docs(mc): describe what the Monte Carlo now scores #565 and docs: bring every surface up to the shipped Monte Carlo and data layer #578: the docs-site multi-agent, agents-api, simulation, architecture and backend-api pages, plus the package READMEs and INDEX. Signature tables checked by introspection.
• The A/B, re-measured through the production path (run_lap(profile='no-llm') plus the CLI's own state builder, toggling only the rivals list): Lusail 1/40 laps changed, Monza 0/44, Marina Bay 17/51. Marina Bay is the case: the legacy layer wanted UNDERCUT and the projection prices the stop at -4.7 positions.

Obsoleted: this issue asked to wire the stop-hazard table as OVERCUT eligibility on the surfaces whose rivals lack is_pitting. /recommend no longer lacks it. Submodule #195 made the endpoint report the flag properly, deriving it from the raw fallback, which is where a pitting rival always arrives since the featured parquet drops pit laps. A measured prior is strictly worse than the fact when the fact is available.

Still open:

1. MCP may still build rivals without is_pitting. Check backend/mcp_tools.py before deciding whether the prior is needed there at all.
2. stop_hazard and gap_density are computed, committed and read by nothing. An adversarial audit of the documentation found this. Either wire them or record in measure_mc_tables.py why they are published without a consumer.
3. Calibration was never started.

[VforVitorio]

Reopening: this was auto-closed by a Closes #557 in an older PR body when v2.1.0 landed, but I had re-scoped it after that with work that is genuinely still open.

Two of the three original parts shipped (the MC docs, and the A/B re-measured through the production path). The third was obsoleted rather than done: it asked to wire the stop-hazard table as OVERCUT eligibility on surfaces whose rivals lack is_pitting, and /recommend no longer lacks it, since submodule #195 made the endpoint report the flag properly.

Still open, and the reason this should not be closed:

1. MCP may still build rivals without is_pitting. Check backend/mcp_tools.py before deciding whether a prior is needed there at all.
2. stop_hazard and gap_density are computed, committed and read by nothing. Either wire them or record in measure_mc_tables.py why they are published without a consumer.
3. Calibration was never started.