feat: replace fixed 80% RAM reserve with bounded-reserve formula

[samarthpatel24]

What

Replace the hardcoded total_ram * 0.80 heuristic with a bounded-reserve function: total - clamp(total * 0.15, 4 GiB, 32 GiB).

• Add estimate_usable_ram(total: int) -> int to hardware/memory.py with inline constants
• Update both call sites (compatibility checker + CLI auto min-params)
• Add parameterized unit tests + update two existing shared-memory APU assertions

Total RAM	Old (20% reserve)	New (bounded)
16 GB	12.8 GB	12.0 GB
32 GB	25.6 GB	27.2 GB
64 GB	51.2 GB	54.4 GB
128 GB	102.4 GB	108.8 GB
1 TB	819.2 GB	992.0 GB

Why

The flat 20% reserve wastes too much on large machines (e.g. 200 GB reserved on a 1 TB box) and doesn't reserve enough on small ones. The bounded formula scales the reserve with machine size while capping it at 32 GiB.

Part 1 of #51

Testing

• Tests pass (pytest) — 321 passed
• New tests added — parameterized tests for 16 GB, 32 GB, 64 GB, 128 GB, 1 TB + boundary checks
• Tested on real hardware — 15.3 GB RAM, RTX 4070 Laptop GPU, ranking works end-to-end

Notes

Part 2 (available-RAM awareness, --planning/--strict-now flags, dual fit tracking) will follow in a separate PR that closes #51.

[samarthpatel24]

Add max(0, ...) guard to match the proposed signature in #51 — prevents negative return if total RAM is somehow below the 4 GiB minimum reserve.

[Andyyyy64]

Reviewed this locally. This is the right first step for #51: replace the fixed 80% RAM rule with a bounded reserve, while leaving current available RAM for the follow-up.

Minor note: estimate_usable_ram has the negative-return guard, but the tests do not cover a total RAM value below the 4 GiB minimum reserve. I do not think that needs to block this PR. The guard is simple, and the normal machine sizes are covered.

I am going to merge this, then answer the Part 2 questions on #51 separately.