Fail model load early when free RAM is below the catalog estimate

[jatinkrmalik]

Summary

Fixes #250. Before WhisperKit/CoreML starts, refuse a load when installed RAM or os_proc_available_memory() is below that model's catalog estimate. The UI shows an error instead of spinning forever under memory pressure.

Test plan

• Unit: SystemInfo.canFitModelInMemory physical / available / unknown / enough cases
• Manual: pick a model larger than free RAM → error toast, spinner clears, previous model stays if one was loaded
• Manual: enough free RAM → load still works

Known limits

• Auto-select (size == nil) skips the gate until a concrete size is known
• Catalog GB is an estimate; peak CoreML use can still be higher
• No load timeout yet if the probe says OK but the engine still hangs

[netlify]

✅ Deploy Preview for voca-mac canceled.

Name	Link
🔨 Latest commit	94c7c65
🔍 Latest deploy log	https://app.netlify.com/projects/voca-mac/deploys/6a9a7285e292d4000861a19c

[greptile-apps]

Greptile Summary

The PR adds an early model-load memory check so known oversized models fail with an error before WhisperKit/CoreML begins loading.

• Compares catalog RAM estimates with installed and currently available memory.
• Preserves an already loaded model when the preflight rejects another load.
• Uses conservative Darwin VM accounting and releases the acquired host send right.
• Adds focused gate, state-preservation, and memory-threshold tests.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
Sources/VocaMac/Models/AppState.swift	Adds the preflight refusal before loading-state mutations and leaves any active model untouched when memory is insufficient.
Sources/VocaMac/Services/SystemInfo.swift	Adds installed- and available-memory capacity checks using free plus inactive Darwin pages with balanced Mach-port cleanup.
Tests/VocaMacTests/AppStateTests.swift	Verifies that rejection avoids a Whisper load and preserves the previously loaded model and selection.
Tests/VocaMacTests/Mocks/MockServices.swift	Makes test model loads deterministic by bypassing host-memory variability unless a test explicitly overrides the gate.
Tests/VocaMacTests/ModelTests.swift	Covers insufficient physical memory, insufficient available memory, unavailable probe data, and sufficient capacity.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Request model load] --> B{Concrete model size known?}
    B -- No --> D[Continue existing load flow]
    B -- Yes --> C{Installed and available memory meet estimate?}
    C -- No --> E[Show error and preserve active model]
    C -- Yes --> D
    D --> F[Start WhisperKit/CoreML load]

Loading

_{Reviews (5): Last reviewed commit: "Use free+inactive only for the low-RAM p..." | Re-trigger Greptile}

[jatinkrmalik]

@greptileai please re-review the current head (low-RAM gate fixes in e47f66a + empty retrigger commit).

[jatinkrmalik]

@greptileai please re-review eaf05b9 — VM accounting now uses free+inactive+purgeable only (no speculative double-count, no compressor).

[jatinkrmalik]

@greptileai please re-review 94c7c65 — available memory is free+inactive only (no speculative/purgeable/compressor overcount).