fix(B9/G4): don't double-count an already-resident model on reload

[alichherawalla]

Problem

makeRoomFor's dirty-ceiling check summed every non-evicted dirty resident and added the incoming spec.sizeMB — but on a reload the incoming replaces its own same-key resident, so that entry is the same RAM. Reloading an already-resident dirty model (e.g. an image model re-selected after a thread change) therefore charged 2× its footprint and was refused for memory even though it already occupied that RAM.

Fix

planEviction already handles this (its alreadyResident rule charges the incoming 0 and never evicts the same-key resident). The separate dirty-ceiling check in makeRoomFor didn't mirror it — now it excludes the incoming's own same-key entry from keptDirtyMB, so the footprint is counted once. The real physics guard is untouched: a fresh dirty model whose own footprint exceeds the balanced ceiling is still refused.

- .filter(r => r.dirtyMemory && !plan.evict.some(e => e.key === r.key))
+ .filter(r => r.dirtyMemory && r.key !== spec.key && !plan.evict.some(e => e.key === r.key))

Test — real manager, red→green

Drives the real modelResidencyManager (registers an already-resident dirty image model, reloads the same key), sized so 1× fits the balanced dirty ceiling but 2× (the double-count) exceeds it:

• reload fits — red before the fix (fits:false, double-counted; verified by reverting), green after;
• a genuinely-oversized dirty model is still refused (no over-correction).

66 residency + 126 adjacent memory tests pass; eslint/tsc/depcruise clean.

On-device note (honest)

B9's precise fault — the double-count crossing the memory ceiling on reload — can't be staged on a physical device: it needs an image model that fits budget at 1× but double-counts past the ceiling. SDXL was the only candidate on the test iPhone, and it's override-only (7.3 GB > 4.9 GB budget, so it bypasses the dirty-ceiling check entirely) and has since been removed. The rigorous proof is the real-manager test above; there's no distinguishing device repro to run — same class as the B5/B8 memory-accounting fixes.

Summary by CodeRabbit

• Bug Fixes

  • Fixed model reload memory calculations so an already-resident model’s dirty memory is not counted twice.
  • Improved handling of dirty models near memory limits, allowing valid reloads while still rejecting genuinely oversized models.

• Tests

  • Added coverage for resident model reloads, oversized dirty models, and platform-state isolation.

[qodo-code-review]

ⓘ Qodo reviews are paused because your trial has ended. Ask your workspace admin to add credits to resume reviews. Manage billing

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: b53cd41a-1a47-45e0-974d-bf37cc0fb3d7

📥 Commits

Reviewing files that changed from the base of the PR and between fe90c63 and 2be6eeb.

📒 Files selected for processing (2)

• __tests__/unit/services/modelResidency.test.ts
• src/services/modelResidency/index.ts

---

📝 Walkthrough

Walkthrough

makeRoomFor now excludes an already-resident entry with the incoming key from retained dirty-memory accounting. Tests cover successful same-key dirty reloads, oversized dirty models, Android memory setup, and platform restoration.

Changes

Reload dirty-memory accounting

Layer / File(s)	Summary
Same-key dirty reload handling src/services/modelResidency/index.ts, __tests__/unit/services/modelResidency.test.ts	makeRoomFor no longer double-counts the dirty footprint of a resident replaced by the incoming model, while tests verify both fitting reloads and rejection of genuinely oversized dirty models.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly matches the main fix: avoiding double-counting an already-resident model on reload.
Description check	✅ Passed	The description covers the problem, fix, tests, and notes, but it omits the template's Type of Change and checklist sections.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/b9-image-model-double-count-reload

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud