context: compaction budget measured on this machine; per-slot ctx from the server (0.4.9)

[mjwsolo]

Two changes, both dynamic on the machine and model, no token literals.

1. Compaction point from measured speed. The supervisor reads llama-server's own print_timing lines from server.log as the session runs (prompt length, prefill t/s, decode t/s per request) and publishes a budget in /status: the largest prompt length at which a full re-read finishes within 60 s and decode keeps at least half of its short-prompt speed, clamped to [n_ctx/4, n_ctx - one step's headroom]. Before measurements the warm-up replay's prefill speed is the prior; without that, the cap (today's behaviour). The fork provider sets it as limit.input, which overflow.ts already prefers. Evidence: on the owner's logs decode halves and prefill drops 4-6x between 16k and 64k prompts, and sessions were reaching 120k before compacting.

2. One source of truth for the context window. After load the supervisor reads the per-slot n_ctx from /props (with -np N the flag is split N ways) and reports it as ctx; the launcher's fixed 32768 fallback is gone. This is the class behind the 'request exceeds the available context size' failures on 24 Sep.

Tests: 11 new unit tests (parsing, rules, prior, clamps, launcher config, supervisor status with a fake /props and a growing log); launcher and warm-up suites pass. Binary rebuilt at fork 86d2d971.

Live verification (isolated eval server, Qwen3.6-35B-A3B Q8, ten-turn task):

• 32k context: ctx read from /props = 32768, budget published at 16384 (measured basis, 13 samples), runtime received it; no compaction needed (max prompt 6.7k).
• Default context: budget moved 97280 -> 73728 as deep samples arrived.
• Forced (16k context, re-read cap 1 s): budget 8192 (floor = 2 x prefix 5919, bounded by the cap); exactly one compaction when the prompt crossed 8192 (max 8433); zero 'exceeds the available context' errors; all ten turns completed.
• Two defects found and fixed by that forced run before shipping: a floor below the fixed prompt prefix compacted on every step; the prefix must come from the first main request, not the smallest prefill.
• Control: Qwen3.8-27B on the same branch reuses the prefix perfectly (24-29 tokens per turn) and passes. Qwen3.6's turn-two full re-read reproduces on main and is model-specific (checkpoint-based reuse), tracked separately.
  Checks 10/10 with the new supervisor and binary.