Add likelihood ratio (LR) calculation endpoints for striation and impression marks

[laurensWe]

This PR implements the LR calculation pipeline for both striation (CCF-based) and impression (CMC-based) mark comparisons, replacing placeholder stubs with real logic.

Core changes:

• conversion/likelihood_ratio.py (new): Loads LR systems from pickle, wraps lir library calls for striation (calculate_lr_striation) and impression (calculate_lr_impression), and provides a ReferenceData
  model holding KM/KNM scores and LLRs. Reference data is currently hardcoded dummy data pending lr_module_scratch integration.
• conversion/utils.py: Adds ccf_score_to_logodds to transform CCF scores before feeding them to the striation LR system.
• processors/controller.py: Implements process_lr_striation and process_lr_impression — full pipeline functions that load marks, compute LR, build result metadata, and save overview plots (CCF/CMC comparison
  overviews with histograms and LLR transformation curves).
• processors/schemas.py: Tightens request schemas — renames mark_ref/mark_comp to mark_dir_ref/mark_dir_comp, flattens CalculateLR to include MetadataParameters directly, adds ImpressionLRParameters
  (JSON-serialisable mirror of ImpressionComparisonMetrics), adds validation (score ≤ n_cells, score range [-1, 1]).
• processors/router.py: Wires up the LR endpoints to the new controller functions; removes TODO stubs; saves aligned marks to disk after score calculation.
• conversion/plots/utils.py (new): build_results_metadata_striation and build_results_metadata_impression helpers that format result metadata dicts for plot display.
• Tests: Replaces monolithic test_schemas.py with split conftest.py + test_schemas.py + test_router.py; adds integration-level coverage for LR endpoints.

[SimoneAriens]

see comments, some parts can be imported from lir or lr_module_scratch and some should be moved to conversion

[SimoneAriens]

almost there, mostly cosmetic changes

[vergep]

• We need to be careful whether LR_system.apply needs score or transformed_score.
  For the rest minor comments.

[github-actions]

Diff Coverage

Diff: origin/main..HEAD, staged and unstaged changes

• packages/scratch-core/src/conversion/likelihood_ratio.py (92.9%): Missing lines 41,43,60
• packages/scratch-core/src/conversion/plots/data_formats.py (100%)
• packages/scratch-core/src/conversion/plots/utils.py (83.3%): Missing lines 398,403-404
• packages/scratch-core/src/conversion/utils.py (100%)
• src/extractors/schemas.py (100%)
• src/processors/controller.py (100%)
• src/processors/router.py (94.7%): Missing lines 67
• src/processors/schemas.py (100%)

Summary

• Total: 160 lines
• Missing: 7 lines
• Coverage: 95%

packages/scratch-core/src/conversion/likelihood_ratio.py

Lines 37-47

  37 
  38     @model_validator(mode="after")
  39     def _validate_matching_lengths(self) -> Self:
  40         if len(self.km_scores) != len(self.km_llrs):
! 41             raise ValueError("km_scores and km_lrs must have the same length")
  42         if len(self.knm_scores) != len(self.knm_llrs):
! 43             raise ValueError("knm_scores and knm_lrs must have the same length")
  44         return self
  45 
  46     @property
  47     def scores(self) -> np.ndarray:

---

Lines 56-64

  56     @property
  57     def llr_intervals(self) -> np.ndarray:
  58         """Concatenated KM and KNM LLR intervals, shape (n, 2)."""
  59         if self.km_llr_intervals is None or self.knm_llr_intervals is None:
! 60             raise ValueError("Only models with llr_intervals can be used")
  61         return np.concatenate([self.km_llr_intervals, self.knm_llr_intervals], axis=0)
  62 
  63     @property
  64     def labels(self) -> np.ndarray:

---

packages/scratch-core/src/conversion/plots/utils.py

Lines 394-408

  394 
  395 def _format_lr(llr_data: LLRData) -> str:
  396     """Format a single log-LR value with optional confidence interval."""
  397     if len(llr_data.llrs) > 1:
! 398         raise ValueError(f"expected single LR value, got {len(llr_data.llrs)}")
  399 
  400     log_lr = llr_data.llrs[0]
  401 
  402     if llr_data.llr_intervals is not None:
! 403         lower, upper = llr_data.llr_intervals[0, 0], llr_data.llr_intervals[0, 1]
! 404         return f"{log_lr:.2f} ({lower:.2f}, {upper:.2f})"
  405     return f"{log_lr:.2f}"
  406 
  407 
  408 def _common_results_metadata(

---

src/processors/router.py

Lines 63-71

  63     include_in_schema=False,
  64 )
  65 async def calculate_score_impression(impression: CalculateScore) -> ComparisonResponseImpression:
  66     """Compare two impression profiles."""
! 67     vault = create_vault(impression.tag)
  68     return ComparisonResponseImpression.generate_urls(vault.access_url)
  69 
  70 
  71 @processors.post(

---

[github-actions]

Package	Line Rate	Branch Rate	Health
.	96%	88%	✔
comparators	100%	100%	✔
computations	100%	100%	✔
container_models	99%	100%	✔
conversion	97%	82%	✔
conversion.export	100%	100%	✔
conversion.filter	97%	89%	✔
conversion.leveling	100%	100%	✔
conversion.leveling.solver	100%	75%	✔
conversion.plots	98%	88%	✔
conversion.preprocess_impression	99%	91%	✔
conversion.preprocess_striation	89%	58%	✔
conversion.profile_correlator	96%	82%	✔
conversion.surface_comparison	93%	85%	✔
extractors	99%	75%	✔
mutations	100%	100%	✔
parsers	98%	80%	✔
parsers.patches	89%	60%	✔
preprocessors	99%	89%	✔
processors	99%	100%	✔
renders	98%	67%	✔
utils	92%	75%	✔
Summary	97% (3058 / 3145)	84% (330 / 394)	✔

Minimum allowed line rate is 50%