fix: normalize activity and workout detail semantics

[ericlitman]

Summary

Tonal activity and workout-detail responses now tell callers what the upstream
data actually supports: strength workouts are separated from external
activities, fixed-page history gaps remain explicitly uncertain, and
unperformed sets no longer look like completed zero-difficulty work.

The CLI and MCP views now agree on user-facing load and performance semantics,
including bilateral StraightBar totals, off-machine duration movements, warm-up
volume, raw-versus-formatted side splits, and structured detail failures.

Behavioral contract

• History limits are enforced locally over Tonal's fixed 50-row activity page,
  with activity type, strength eligibility, and completeness metadata exposed
  instead of an unverified pagination promise.
• Raw detail excludes Rest rows and distinguishes resistance work from
  off-machine duration movements without making mixed-workout totals unsafe to
  sum.
• An upstream execution sentinel determines whether difficulty, consistency,
  and suggested-weight zeroes are real measurements or placeholders from a set
  that never started. Legitimate zero scores and configured mode flags remain
  intact.
• Bilateral StraightBar load, 1RM, and suggestion fields use one total-load
  convention across CLI detail, MCP detail, and MCP exercise history; upstream
  volume is not doubled.
• Detail 404s distinguish known external activities, temporarily unavailable
  strength detail, definitive absence on an exhausted page, and uncertainty
  beyond a full page.

Validation

• python3 -m unittest discover -s tests -v — 25 tests passed.
• python3 -m compileall -q mcp tonal_tool.py tests — passed.
• git diff --check — passed.
• A live reference workout containing one performed set and three never-started
  sets returned null analytics only for the unperformed rows; the CLI
  aggregate changed from the fabricated 0.06 to the performed-only 0.23.
• Compound review findings were fixed and regression-tested; the final
  read-only review-agent pass returned no findings.

Post-Deploy Monitoring & Validation

• Window and owner: repository maintainer, first 24 hours after merge and the
  next live Tonal detail invocation.
• Smoke test: run CLI detail <activity_id> and MCP get_workout_detail for a
  workout containing performed and skipped sets; confirm skipped analytics are
  null, performed zero scores remain 0, and mode flags remain present.
• Cross-surface check: compare a bilateral StraightBar set in CLI detail, MCP
  detail, and MCP exercise history; load and 1RM must agree while volume stays
  unchanged.
• Error-path check: request detail for a known External activity and an unknown
  ID; expect no_strength_data, or activity_type_unknown when the observed
  50-row page cannot prove absence.
• Failure signals: TypeError, unsupported operand type, raw placeholder-zero
  analytics entering averages, divergent StraightBar loads, or a definitive
  activity_not_found result paired with a full source page.
• Telemetry: this repository has no centralized runtime dashboard. Inspect the
  invoking CLI/MCP process logs for the failure strings above and watch GitHub
  issues for response-shape regressions during the validation window.
• Rollback trigger: any exception on previously readable workout detail or a
  reproducible CLI/MCP load mismatch. Revert this PR and retain the captured
  response fixture before retrying the rollout.

Fixes TONAL-6.

---