Add a self-supervised window-summary head and the window-statistic probes

[amrit110]

Frozen probes on full_run_v10 (scripts/probe_summary_signal.py, new) show
the backbone's state holds current levels of the GBM's signal panel but
not changes from the visit's first value (creatinine delta R^2 0.00) and
not fine window counts, under a linear and a tree probe alike. Those are
the contents of the GBM's margin on AKI and the counting events.

This adds an auxiliary, training-only objective that asks the state to
report those statistics:

• odyssey/training/summary_targets.py: the target panel (per-signal
  min/max/mean over 6 h and 24 h, change from visit-first, per drug-class
  and per-family counts over 6 h and 24 h), computed by the GBM's own
  feature code at 4-hourly landmark rows; standardization (counts through
  log1p, winsorized at 0.5/99.5 percentiles); per-visit tables; and the
  per-chunk lookup that places each target on the last token of its
  bundle. Targets are computed from the input and never fed in.
• odyssey/models/summary_head.py: a linear (or MLP) head and a masked
  Huber loss that is zero with a live graph where no target applies.
• Both sequence models take summary_targets/summary_head_hidden, add the
  weighted loss to the streaming and steering objectives, and report a
  summary_loss component.
• TrainingConfig gains summary_targets_dir, summary_weight,
  summary_head_hidden and summary_num_targets (recorded at training time
  so a checkpoint rebuilds the same head); the training loop, validation
  and load_run are wired accordingly. None (the default) reproduces every
  existing run.
• scripts/build_summary_targets.py precomputes the targets per shard in
  parallel and fits the stats.

Tests cover the target names, landmark rows, target values (baseline
change, counts, NaN for unmeasured), stats, JSON round trip, chunk
placement on bundle ends, padding, the head, the masked loss, and the
model wiring. 1927 tests pass; ruff, format and mypy clean.

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01StsjDmueuDqVNNuVoCEF1P

[codecov]

Codecov Report

❌ Patch coverage is 96.03175% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.12%. Comparing base (31144a5) to head (3474f2c).

Files with missing lines	Patch %	Lines
odyssey/training/train.py	80.00%	6 Missing ⚠️
odyssey/training/summary_targets.py	98.32%	3 Missing ⚠️
odyssey/inference/run_inference.py	83.33%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #251      +/-   ##
==========================================
+ Coverage   89.98%   90.12%   +0.13%     
==========================================
  Files          69       71       +2     
  Lines       10771    11023     +252     
==========================================
+ Hits         9692     9934     +242     
- Misses       1079     1089      +10     

Files with missing lines	Coverage Δ
odyssey/models/sequence_model.py	96.45% <100.00%> (+0.18%)	⬆️
odyssey/models/summary_head.py	100.00% <100.00%> (ø)
odyssey/inference/run_inference.py	93.81% <83.33%> (-0.15%)	⬇️
odyssey/training/summary_targets.py	98.32% <98.32%> (ø)
odyssey/training/train.py	88.28% <80.00%> (-0.40%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.