revert(mcp): back out leadbay_getting_started (#175)

[ArtyETH06]

Backs out the #175 merge (24cf7cd). main returns to e4a8280's parent state.

#175 was merged at 13:43 UTC; Codex's review landed at 13:49 — six minutes later, so its findings never gated the merge. Eight are open, four of them P1:

• Paid reveal can overspend. Gate 4 tells the agent to pass a singular leadId, but leadbay_enrich_titles only reads leadIds?: string[]. The unknown key is ignored, so the confirmed call falls back to the default wishlist selection with confirm/email set — it can spend on the whole wishlist instead of the one drafted lead.
• Manifest returned by reference. The server attaches _meta to result objects in place; those stick to the module-level singleton and leak into later calls.
• Eval evidence is not trustworthy. runSessionLive records every call as ok: true, output_len: 0 and never updates from tool_result, so a failed call still satisfies required_calls and reaches the judge as a success.
• The overdeliver scenario can spend real quota. It relies on backendFixtures to block the paid launch, but the live runner ignores fixtures.

Plus four P2s: the tour dead-ends at gate 4 under LEADBAY_MCP_WRITE=0 (enrich_titles is a write tool); WORKFLOWS.md still requires the CRM/scheduling gates that were removed; the STOP include forces a final line that contradicts "the 1:1 offer must be last"; allowed_calls is declared but never enforced.

Nothing shipped — auto-tag.yml only fires on packages/mcp/package.json, which #175 never touched. npm latest is still 0.28.0 from 717b02c.

Gates on this branch: prompts:check clean, build, typecheck, and 1,289 tests all pass.

Supersedes #176 and #177, both closed — one decision should be one PR. The re-land with the eight fixes comes back as its own PR once this lands.