`--diff` smart mode over-weights directory proximity, under-weights import graph

[nikolay-e]

Problem

Smart mode (--diff) includes irrelevant sibling files from the same directory and misses direct import dependencies of changed files. It also over-selects chunks from tangentially related files.

Observed behavior

• Irrelevant sibling files included: InlineLyricsPanel, LyricsView, SeekBar, SleepTimerModal, VolumeControls, useAlbumColors — zero connection to the diff
• Direct dependencies missed: FavoriteButton.tsx and useFavorites.ts — newly imported in ArtistDetailPage.tsx but not selected
• Over-selection of chunks: player.store.ts pulled in almost entirely (4 chunks, ~800 lines) when only ~10 lines related to sendSignal are relevant
• Token budget: smart mode produces 28K tokens vs 10K for --full, with an ideal target of ~12-15K

Root cause hypothesis

Scoring over-weights directory proximity (sibling files) and under-weights import graph proximity. Files in the same folder get high scores regardless of actual semantic relevance to the diff, while direct imports added in the diff are not sufficiently boosted.

Reproduction

• Repository: nikolay-e/yay-tsa
• Commits: 0fe7132..119b0f0
• TreeMapper version: 1.4.3

treemapper . --diff 0fe7132..119b0f0

Expected behavior

Smart mode should:

1. Prioritize files connected via import graph (direct dependencies of changed files)
2. Reduce weight of directory proximity when there is no import/reference relationship
3. Select only relevant chunks from large files (e.g. ~10 lines from player.store.ts, not ~800)
4. Stay within a reasonable token budget (~12-15K for this diff)

[nikolay-e]

Status Update (v1.4.4+)

Retested with current main (commit f95022d) on the original reproduction case (0fe7132..119b0f0 in nikolay-e/yay-tsa).

What's fixed

• Irrelevant sibling files eliminated: InlineLyricsPanel, LyricsView, SeekBar, SleepTimerModal, VolumeControls, useAlbumColors — none included in precise or auto modes
• Over-selection fixed: player.store.ts no longer pulled in (was 4 chunks / ~800 lines)
• Token budget: 39 fragments from 4 relevant files (down from 28K tokens)

What remains

• Direct import dependencies not selected: FavoriteButton.tsx and useFavorites.ts (newly imported in ArtistDetailPage.tsx) are still not included in any scoring mode. Root cause: the utility/needs system generates information needs from diff text identifiers only — newly added imports don't create matching needs, so greedy selection assigns zero marginal gain even when PPR scores are high.

Mode comparison on this diff

Mode	Fragments	Files	Noise siblings	Import deps
precise (PPR)	39	4	✅ excluded	❌ missed
discover (EgoGraph)	128	22	❌ all included	❌ missed
auto	39	4	✅ excluded	❌ missed

The import dependency issue is tracked separately — it requires extending the needs system to include identifiers from graph neighbors of core fragments, not just diff text.