feat(training-history): date filtering, pagination, and detail tiers

[dcaslin]

Problem

get_training_history returns the user's entire workout history in one JSON
blob. On a representative account that's 231 workouts / ~3.4 MB, of which
98.9% is the per-set records data (sets average ~14.6 KB/workout, dominated
by redundant fields like previous_*, archived_*, valueEmpty). MCP clients
(e.g. Claude Desktop) reject it with "result too large", so the tool is
effectively unusable on any account with real history.

The upstream /programs/history endpoint accepts only timezone_offset and
returns everything grouped by date — there are no server-side filtering or
pagination knobs. So all shaping is done client-side in the MCP server.

What changed

get_training_history gains optional parameters (all backward-compatible — a
bare call still works, now bounded and useful):

Param	Purpose
start_date / end_date	Inclusive YYYY-MM-DD date-range filter
detail	summary (default) or full
page / page_size	Walk large histories in chunks (has_more flag)

• summary (default): per workout — date, title, program, week/day,
  exercise_count, exercise names, and computed total_volume. ~500 bytes each.
• full: adds records with each set slimmed to the meaningful fields
  (weight, reps, target, rpe, skipped, weight_unit), dropping the
  redundant bloat. Roughly halves full-mode size with no analytical loss.
• Guardrail: page_size is capped per tier (summary ≤ 100, full ≤ 25) so no
  single response can blow past the client limit regardless of arguments.
• Envelope: every response carries pagination (incl. has_more),
  filters, and a hint telling the model how to fetch more.

A bare get_training_history() now returns the summary of the 50 most recent
workouts instead of the full 3.4 MB dump.

Architecture

All transforms live in a new pure module src/boostcamp_mcp/history.py
(flatten → date-filter → newest-first sort → detail-shape → paginate →
envelope). server.py validates params, calls the API once, delegates to
history.shape_history, and returns JSON. The pure module is fully unit-tested
with no network.

Test plan

• uv run pytest — 22 unit tests covering set slimming, volume math
  (skipped/empty sets), summary/full shaping, date-filter boundaries,
  newest-first ordering, pagination edges, page_size caps, and validation.
• Live smoke test against a real account:
  • default summary (5 workouts): 2.5 KB response (was 3.4 MB), has_more: true
  • detail="full" (3 workouts): ~12 KB with slimmed sets
  • start_date="2026-04-07": 20 workouts, correct filtering
  • bad date → clean Error: start_date must be YYYY-MM-DD message
• Server object loads cleanly under FastMCP.

Design and implementation notes: docs/superpowers/specs/2026-06-12-training-history-pagination-design.md.

🤖 Generated with Claude Code