fix(triage_calibration): enum mismatch — expects small/large, SODA uses low/high

[decko]

Summary

TriageCalibrationMetric silently skips sessions where triage complexity is low or high because the filter checks for small/medium/large:

if complexity not in ("small", "medium", "large"):
    continue

SODA's triage schema uses low/medium/high:

Complexity string `json:"complexity" jsonschema:"enum=low|medium|high"`

Only medium sessions qualify, artificially deflating the metric. In v0.7.0 (n=6), 4 sessions were excluded — reported calibration was 0.50 (1/2) when the actual calibration is 0.83 (5/6).

Fix

Accept both enum sets in triage_calibration.py:

# Map SODA enum to calibration bands
COMPLEXITY_MAP = {
    "small": "small", "low": "small",
    "medium": "medium",
    "large": "large", "high": "large",
}

Affected file

src/raki/metrics/operational/triage_calibration.py

Impact

All historical SODA sessions with low or high complexity have been excluded from calibration scoring. Fixing this will retroactively correct all past reports.

[decko]

Spec

Problem

TriageCalibrationMetric silently excludes sessions where SODA reports complexity as low or high, because the filter at src/raki/metrics/operational/triage_calibration.py:85 checks for ("small", "medium", "large") — the raki-internal enum — while SODA's triage schema emits low/medium/high.

Only medium sessions pass the filter. In practice this means 4 out of 6 sessions in a v0.7.0 run were excluded, producing a reported calibration of 0.50 (1/2) when the correct figure is 0.83 (5/6).

Fix

Introduce a COMPLEXITY_MAP that normalises both enum sets to the internal small/medium/large bands before filtering and scoring:

COMPLEXITY_MAP: dict[str, str] = {
    "small": "small",
    "low": "small",
    "medium": "medium",
    "large": "large",
    "high": "large",
}

Replace the raw complexity lookup with COMPLEXITY_MAP.get(complexity) throughout the compute() method. Sessions whose complexity field is not in the map are still skipped (unchanged behavior for unknown values).

Update the docstring and description / interpretation strings to document both enum variants.

Acceptance Criteria

• COMPLEXITY_MAP defined at module level in triage_calibration.py
• Sessions with complexity="low" are treated as "small" band
• Sessions with complexity="high" are treated as "large" band
• Sessions with complexity="medium" behavior unchanged
• Sessions with unmapped complexity values still skipped (no change)
• per_complexity dict keys remain "small", "medium", "large" (no downstream breakage)
• Unit tests cover: low → small, high → large, medium unchanged, unknown skipped
• Docstring updated to mention both SODA and raki enum variants
• Towncrier fragment 346.fix

---

Plan

Step 1 — Add COMPLEXITY_MAP and update compute() in triage_calibration.py

File: src/raki/metrics/operational/triage_calibration.py

1. Add after the cost threshold constants:

# Normalise SODA enum (low/medium/high) to internal bands (small/medium/large).
COMPLEXITY_MAP: dict[str, str] = {
    "small": "small",
    "low": "small",
    "medium": "medium",
    "large": "large",
    "high": "large",
}

2. In compute(), replace:

complexity = output.get("complexity")
if complexity not in ("small", "medium", "large"):
    continue

with:

raw_complexity = output.get("complexity")
complexity = COMPLEXITY_MAP.get(raw_complexity or "")
if complexity is None:
    continue

All subsequent references to complexity (lines 95, 97, 106) remain unchanged — they already use the normalised band name.

3. Update the description string to read: "…Only sessions with a triage complexity estimate (small/medium/large or SODA equivalents low/high) and an actual cost are included."

Step 2 — Add unit tests

File: tests/test_triage_calibration.py (or nearest existing file for this metric)

Add parametrized cases:

• complexity="low" with cost within small threshold → calibrated (score contribution 1.0)
• complexity="high" with any cost → calibrated (large band, no upper bound)
• complexity="medium" unchanged behavior
• complexity="unknown" → skipped (not counted in total)
• Mixed dataset: ["low", "medium", "high", "unknown"] → total=3, correct mean

Step 3 — Towncrier fragment

File: changes/346.fix

``TriageCalibrationMetric`` now accepts SODA's ``low``/``high`` complexity enum variants in addition to the raki-native ``small``/``large`` bands. Previously, sessions with ``low`` or ``high`` complexity were silently excluded, causing artificially deflated calibration scores.

Estimated scope

• ~20 lines changed in triage_calibration.py
• ~25 lines of new tests
• 1 changelog fragment