fix(meta_ads): count conversions via a canonical exact-match counter (#340)

[hyoshi]

Summary

meta_row_conversions (the Meta conversion counter feeding CPA → anomaly
detection / diagnose_performance / budget-efficiency) counted via a
substring scan "lead" in action_type or "purchase" in action_type, which
over-counted for two reasons:

1. Alias double-count. Meta's Insights actions array returns one
   conversion under several overlapping action_type rows. The generic lead
   is, by Meta's own definition, the aggregate ("all offsite + all
   On-Facebook leads") of its components offsite_conversion.fb_pixel_lead +
   onsite_conversion.lead_grouped; purchase/omni_purchase is the
   analogous roll-up. Because actions is fetched unfiltered, aggregate +
   component rows co-occur, so summing every *lead*/*purchase* row counted
   the same conversions two or three times.
2. Substring false positives. Operator-named custom conversions
   (offsite_conversion.custom.<slug>) carry free-text slugs that may contain
   lead/purchase and were swept in.

This deflated the absolute CPA emitted by diagnose_performance and the
budget-efficiency scorer (the anomaly-detector CPA-spike path is ratio-based
and largely shielded). It also diverged from the MCP-analysis counter
_extract_cv, which already used an exact {lead, purchase, complete_registration} set — so multiple live counters disagreed on what
counts as a conversion.

Fix

A single canonical counter mureo/meta_ads/_conversion_count.py
(count_conversions_from_actions + CONVERSION_ACTION_TYPES) that counts
only the deduped generic conversion action_types — the aggregate already
includes its components, so components are not added on top, and custom slugs
no longer match. All four live counters now route through it so they cannot
drift again:

• mureo/analytics/builtin/_common.py meta_row_conversions (lazy import to
  keep adapter registration free of the mureo.meta_ads client weight; BYOD
  branch unchanged)
• mureo/meta_ads/_analysis.py _extract_cv (delegates)
• mureo/analytics/builtin/_budget_efficiency.py
• mureo/meta_ads/_insights.py breakdown

The campaign-type classifier (_classify_result_indicator) keeps substring
matching — it is a boolean "any conversion signal?" check, not a counter, so
substring is correct there.

Behaviour notes

• complete_registration is now counted on the analytics/CPA path (previously
  only _extract_cv counted it) — an intentional consistency change.
• Known component-only / custom-event under-count (a family reported only
  as a component row, or as a custom event) is tracked separately as the
  operator-canonical-event enhancement (feat(meta_ads): operator-specified canonical conversion event (avoid component-only under-count) #342).

Test plan

• tests/test_meta_conversion_count.py — dedup (aggregate+component),
  custom-slug false positive, view/engagement exclusion, non-list/empty/junk
  handling, _extract_cv byte-for-byte equivalence.
• Updated the two fixtures that encoded the old substring behaviour
  (test_live_clients.py, test_budget_efficiency.py) to realistic
  aggregate+component arrays that prove the dedup; added a budget-efficiency
  custom-slug regression test.
• ruff + black clean on mureo/; mypy --strict clean on changed
  modules; 1125 passed in the meta/analytics sweep (only the known
  installed-plugin env-gating noise excluded).
• Reviewed by python-reviewer (adversarial) — it caught two further
  substring counters (_budget_efficiency.py, _insights.py) that this PR now
  also migrates.

Closes #340

https://claude.ai/code/session_011rAu94b1o1xWYZhARk1VmL