Surface per-line OCR confidence so the low-confidence filter runs

[FuJacob]

Summary

ScreenTextExtractor discarded Vision's per-observation confidence, so OCRTextHygiene.dropLowConfidence — the cheapest, highest-signal hygiene filter — only ever saw a synthesized 1.0 and dropped nothing. This carries each recognized line's real confidence through a new ExtractedScreenText.lines field and feeds those lines to OCRTextHygiene.clean, so the recognizer's weakest guesses are dropped before they reach the completion prompt.

Validation

xcodebuild -project Cotabby.xcodeproj -scheme Cotabby -destination 'platform=macOS' \
  build-for-testing -derivedDataPath build/DerivedData
# ** TEST BUILD SUCCEEDED **

swiftlint lint --quiet
# 0 violations

• New test test_generateContext_dropsLowConfidenceOCRLines proves the wiring end-to-end: a clean, plausible sentence at 0.2 confidence is dropped while a 0.95-confidence line survives, even though no other hygiene filter would catch the former.
• The existing generator tests are updated to populate lines (at a confidence above the threshold), so they keep exercising the symbol/digit/sanitizer filters unchanged.

Linked issues

None.

Risk / rollout notes

• Behavior change (visual context only). Real OCR confidence now gates lines via the existing dropLowConfidence threshold (0.4). Lines Vision is unsure about no longer enter the prompt; this only affects the screen-context excerpt, not completion generation.
• ExtractedScreenText.lines defaults to empty for any caller that supplies only joined text, so the type stays backward compatible.
• OCRTextHygiene.OCRLine gains Sendable (pure value type) to cross the OCR continuation boundary inside the Sendable ExtractedScreenText.

Greptile Summary

This PR threads Vision's per-observation confidence through ExtractedScreenText.lines so OCRTextHygiene.dropLowConfidence — previously neutralized by a synthesized 1.0 — now filters on real recognizer scores before lines reach the completion prompt.

• ExtractedScreenText gains a lines: [OCRTextHygiene.OCRLine] field (default-empty, backward compatible), and ScreenTextExtractor populates it with each top candidate's actual confidence value.
• ScreenshotContextGenerator replaces the old re-split-with-fake-confidence pattern with a direct pass of extracted.lines to OCRTextHygiene.clean.
• OCRLine picks up Sendable conformance, and a new end-to-end test confirms the wiring: a 0.2-confidence line is dropped while a 0.95-confidence line survives.

Confidence Score: 4/5

Safe to merge — the core wiring is correct and the new test provides solid end-to-end coverage of the confidence gate.

The only finding is a one-sentence doc-comment in OCRTextHygiene.swift that still says the extractor currently discards confidence, directly contradicting what this PR accomplishes. It does not affect runtime behavior.

Cotabby/Support/OCRTextHygiene.swift — the OCRLine doc-comment contains a now-false claim about confidence being discarded.

Important Files Changed

Filename	Overview
Cotabby/Services/Visual/ScreenTextExtractor.swift	Adds lines: [OCRTextHygiene.OCRLine] to ExtractedScreenText (default-empty for backward compatibility) and rewires the Vision callback to capture candidate.confidence per observation instead of discarding it.
Cotabby/Services/Visual/ScreenshotContextGenerator.swift	Switches from extracting only .text (then re-splitting with synthetic confidence 1.0) to storing the full ExtractedScreenText and passing extracted.lines directly to OCRTextHygiene.clean, enabling real confidence gating.
Cotabby/Support/OCRTextHygiene.swift	Adds Sendable conformance to OCRLine (correct). The doc-comment on OCRLine.confidence contains a stale sentence claiming the extractor currently discards confidence, which this PR directly contradicts.
CotabbyTests/ScreenshotContextGeneratorTests.swift	Existing helpers updated to supply OCRLine objects at confidence 0.9 to preserve prior test coverage. New test proves end-to-end wiring: 0.2-confidence line is dropped, 0.95-confidence line survives.

Sequence Diagram

sequenceDiagram
    participant SCG as ScreenshotContextGenerator
    participant STE as ScreenTextExtractor
    participant Vision as Vision OCR
    participant Hyg as OCRTextHygiene.clean

    SCG->>STE: extractText(from: image)
    STE->>Vision: VNRecognizeTextRequest
    Vision-->>STE: [VNRecognizedTextObservation]
    Note over STE: Build OCRLine(text, confidence)<br/>for each top candidate
    STE-->>SCG: ExtractedScreenText(text, lineCount, lines)
    Note over SCG: lines carry real Vision confidence<br/>(was: re-split text with confidence=1.0)
    SCG->>Hyg: clean(lines: extracted.lines, maxChars:)
    Note over Hyg: 1. dropLowConfidence (threshold 0.4) now effective<br/>2. dropReplacementCharacter<br/>3. dropHighSymbolDensity<br/>4. dropDigitSubstitution<br/>5. dropLowWordCharacterRatio<br/>6. strip fieldText echoes
    Hyg-->>SCG: cleaned OCR string
    SCG-->>SCG: return VisualContextExcerpt

Loading

_{Reviews (1): Last reviewed commit: "Surface per-line OCR confidence so the l..." | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.

[greptile-apps]

The doc-comment on OCRLine.confidence says "The orchestrating extractor currently discards Vision's per-candidate confidence" — but that statement is precisely what this PR fixes. Leaving it in place makes the comment actively misleading for anyone reading the type definition after the change lands.

Suggested change

	/// Confidence is carried alongside the text because the cheapest, highest-signal filter
	/// (`dropLowConfidence`) needs it. The orchestrating extractor currently discards Vision's
	/// per-candidate confidence; surfacing it into this value type is what lets filter #1 run.
	struct OCRLine: Equatable {
	struct OCRLine: Equatable, Sendable {
	/// Confidence is carried alongside the text because the cheapest, highest-signal filter
	/// (`dropLowConfidence`) needs it. `ScreenTextExtractor` surfaces Vision's per-candidate
	/// confidence here so filter #1 operates on real recognizer scores instead of a constant.
	struct OCRLine: Equatable, Sendable {

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!