Add clipboard relevance filter and line-level distillation

[FuJacob]

Summary

Clipboard content was blindly injected into every autocomplete prompt, even when stale
or noisy. This PR adds two layers of filtering:

1. Relevance filter (ClipboardRelevanceFilter) — gates clipboard injection on three
   zero-latency heuristics: staleness (>5 min unchanged → drop), app affinity (same-app
   copy always passes), and token overlap (no shared words with prefix → drop).

2. Content distillation (ClipboardContentDistiller) — for clipboard that passes the
   relevance gate, extracts only the lines whose tokens overlap with the user's prefix
   text. Short clipboard (≤3 lines) passes through unchanged. When no individual line
   overlaps, the first 300 characters are kept as a head-biased fallback.

Also extracts shared tokenization logic into PromptContextSanitizer.significantTokens()
so both components use the same tokenizer.

Validation

xcodebuild -project tabby.xcodeproj -scheme tabby -destination 'platform=macOS' build
# ** BUILD SUCCEEDED **

xcodebuild -project tabby.xcodeproj -scheme tabby -destination 'platform=macOS' build-for-testing
# ** TEST BUILD SUCCEEDED **

swiftlint lint --quiet
# No new warnings

Linked issues

None.

Risk / rollout notes

• Users with clipboard context enabled will see filtered/distilled clipboard instead of raw. This is purely additive filtering — content that was previously injected may now be dropped or trimmed, but nothing new is added.
• The ClipboardContextProviding protocol gains a currentChangeCount property. Any test doubles conforming to this protocol will need updating.
• Tuning constants: staleness threshold (5 min), minimum token length (3 chars), compact line threshold (3 lines), head fallback (300 chars) — all static let on their respective types.

Greptile Summary

This PR adds two layers of clipboard filtering to the autocomplete pipeline: a ClipboardRelevanceFilter that gates injection on staleness and token overlap, and a ClipboardContentDistiller that trims long clipboard to only the lines that share tokens with the user's prefix. It also addresses three previously identified bugs — the boot-time staleness clock reset, the erroneous app-affinity bypass, and the filter/distiller window mismatch — by adopting a nil-sentinel baseline, removing app-affinity logic entirely, and computing the truncated prefix once in the coordinator before passing it to both components.

• ClipboardRelevanceFilter: state machine keyed on NSPasteboard.changeCount; first observation silently records the baseline without starting the staleness clock, so content copied before Cotabby launched is never injected until the user performs a fresh copy.
• ClipboardContentDistiller: clips short (≤3 line) content through unchanged; longer content is filtered to token-overlapping lines with a 300-char head fallback when no line matches.
• SuggestionRequestFactory: truncatedPromptPrefix promoted to internal and prefixText threaded through to the distiller so the filter and distiller evaluate the same bounded window.

Confidence Score: 5/5

Safe to merge. The change is purely additive filtering — clipboard content that was previously injected may now be dropped or trimmed, but no new content is introduced into prompts.

All three bugs flagged in prior review rounds have been addressed: the nil-sentinel baseline prevents the boot-time clock reset, app-affinity logic is removed entirely, and the coordinator now computes the truncated prefix once and passes it to both the filter and the distiller. New types are @MainActor-isolated, injected via narrow protocols consistent with AGENTS.md conventions, and covered by deterministic unit tests with a swappable dateProvider.

The project.pbxproj has a LlamaSwift in Frameworks build file entry that appears unrelated to this feature and has no matching Frameworks build phase addition in the diff — worth confirming it is intentional.

Important Files Changed

Filename	Overview
Cotabby/Support/ClipboardRelevanceFilter.swift	New filter type addressing three previously filed bugs: removes false app-affinity bypass, uses nil-sentinel baseline to avoid the boot-time staleness clock reset, and evaluates token overlap against the same bounded window the distiller sees. Logic, state transitions, and test coverage all look correct.
Cotabby/Support/ClipboardContentDistiller.swift	New distiller correctly passes short (≤3 line) clipboard through unchanged, filters longer content to token-overlapping lines, and falls back to the first 300 chars when no line matches. Uses the shared significantTokens tokenizer for consistency.
Cotabby/App/Coordinators/SuggestionCoordinator+Prediction.swift	Computes the truncated prefix once and passes it to the relevance filter so the filter and the downstream distiller evaluate overlap against the same bounded window. Correctly replaces the old unconditional clipboard injection.
Cotabby/Support/SuggestionRequestFactory.swift	Promotes truncatedPromptPrefix from private to internal so the coordinator can use it, threads prefixText into activeClipboardContext, and adds the ClipboardContentDistiller call before the existing character cap. Changes are minimal and non-breaking.
Cotabby/Models/SuggestionSubsystemContracts.swift	Adds currentChangeCount to ClipboardContextProviding and introduces the ClipboardRelevanceFiltering protocol, both @MainActor-scoped and consistent with existing contract style.
CotabbyTests/ClipboardRelevanceFilterTests.swift	Good coverage of nil input, baseline gating, token overlap (case-insensitive, short-token exclusion), staleness expiry, and clock reset on new copy. Uses injectable dateProvider to make time-dependent assertions deterministic.
CotabbyTests/ClipboardContentDistillerTests.swift	Covers the short-passthrough, partial-overlap, no-overlap head-fallback, case-insensitive, short-token, and empty-prefix cases. All assertions match the documented behaviour.

Sequence Diagram

sequenceDiagram
    participant SC as SuggestionCoordinator
    participant CCP as ClipboardContextProvider
    participant SRF as SuggestionRequestFactory
    participant CRF as ClipboardRelevanceFilter
    participant CCD as ClipboardContentDistiller
    participant PSan as PromptContextSanitizer

    SC->>CCP: currentContext()
    CCP-->>SC: rawClipboard (String?)
    SC->>CCP: currentChangeCount
    CCP-->>SC: pasteboardChangeCount (Int)
    SC->>SRF: truncatedPromptPrefix(from: rawContext.precedingText)
    SRF-->>SC: truncatedPrefix
    SC->>CRF: filter(clipboard:, pasteboardChangeCount:, precedingText: truncatedPrefix)
    note over CRF: 1st call → record baseline, return nil
    CRF->>PSan: significantTokens(from: clipboard)
    CRF->>PSan: significantTokens(from: precedingText)
    PSan-->>CRF: "Set<String>"
    CRF-->>SC: clipboardContext (String? — nil if stale or no overlap)
    SC->>SRF: buildRequest(context:, clipboardContext:, ...)
    SRF->>SRF: truncatedPromptPrefix(from: context.precedingText)
    SRF->>PSan: sanitize(rawContext)
    PSan-->>SRF: sanitizedContext
    SRF->>CCD: distill(clipboard: sanitizedContext, prefixText:)
    CCD->>PSan: significantTokens(from: prefixText)
    CCD->>PSan: significantTokens(from: each line)
    PSan-->>CCD: "Set<String>"
    CCD-->>SRF: distilled (relevant lines or head fallback)
    SRF->>SRF: clippedText(distilled, max: 1200)
    SRF-->>SC: SuggestionRequestBuildResult

Loading

_{Reviews (6): Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." | Re-trigger Greptile}

Context used:

• Context used - AGENTS.md (source)