[P0][LOCAL-04] Hardware/resource profiling and optimization

Programme: #1
Epic: #21

## Objective
Make local AI usable across realistic laptops/workstations through measured hardware detection, model/context selection and cooperative resource scheduling.

## Scope
- Detect CPU, RAM, accelerator type, usable VRAM/shared memory and storage.
- Hardware cohorts and recommended model/quantization/context profiles.
- Runtime telemetry: cold/warm start, TTFT, task latency, memory, power, thermal throttling and browser/speech responsiveness.
- Explicit cache eviction and model residency policy.
- Battery/thermal-aware background scheduling.
- GPU/CPU priority coordination between coding, speech, browser rendering and UI.
- OOM prediction/recovery and smaller-pack fallback suggestions.

## Product rules
- Installed RAM does not equal usable accelerator memory.
- Speech/stop/direct interaction take priority over optional background reasoning.
- Product support claims are tied to declared reference hardware measurements.

## Acceptance criteria
- [ ] Reference hardware is classified reproducibly and selects a supported profile.
- [ ] Speech and deterministic UI remain usable during declared coding workloads.
- [ ] OOM/thermal pressure degrades gracefully without crash loops or hidden cloud fallback.
- [ ] Cache/model eviction is explicit and bounded.
- [ ] Performance dashboard separates model latency from deterministic UI/connector latency.
- [ ] Battery/thermal policies are testable and user-overridable within safe bounds.

## Dependencies
- LOCAL-01
- CORE-03

**First phase:** P0
**Maturity target:** P5
**Owner:** local-ai-speech

No files, tests, or entry points are named, so start by tracing the LOCAL-01 and CORE-03 dependencies and locating the existing local hardware, runtime, and scheduling code. Done means reproducible hardware profiles, graceful OOM and thermal degradation, explicit bounded eviction, separated latency measurements, and testable user-overridable battery policies.

Written by the indexing model from the issue text.