internal/router: priority queues so batch jobs yield to interactive requests

The GPU is a size-1 resource, so a single long job monopolises the box for its
whole duration and every interactive request queues behind it. Callers can now
declare intent with an X-LlamaSwap-Priority header and the serial scheduler
dispatches by score instead of by arrival.

- X-LlamaSwap-Priority: signed integer, 0 default, absent/unparseable means 0.
  interactive/normal/batch aliases resolve to +100/0/-100. Values are not
  clamped: the caller composes band and any per-user offset itself.
- serial dispatch score = priority + swap affinity + aging. Bands sit 100 apart
  so a small caller offset orders work inside a band without crossing one;
  aging is unbounded so low-priority work cannot starve.
- routing.scheduler.settings.serial.{agingDivisor,swapAffinityBonus}, defaulting
  to 60s/point and +10. swapAffinityBonus is capped at 99 so it can never
  promote a request into the next band.
- fifo adds the header to its per-model priority, so the header is not silently
  ignored under that scheduler.
- /metrics exports per-band queue depth, oldest wait and dispatch counts, plus
  counters for how often aging or swap affinity changed the pick. Each request
  records its priority, band, score and queue wait in the activity log.

Note swapAffinityBonus defaults to 10, so equal-priority requests for the
already-loaded model now run before older requests that need a swap. Set it to
0 for the previous strict arrival order.

fixes #9

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01WUyhZBgv8BBCC5MduX88gE

config-schema.json

@@ -605,17 +605,37 @@
                                 "fifo"
                             ],
                             "default": "serial",
-                            "description": "Scheduler to use. 'serial' (default on this fork): strict one-model-at-a-time, requests run in exact arrival order, switching models evicts every other model first. 'fifo': throughput-oriented, batches same-model requests and allows parallel/co-resident models."
+                            "description": "Scheduler to use. 'serial' (default on this fork): strict one-model-at-a-time, highest-priority request runs next, switching models evicts every other model first. 'fifo': throughput-oriented, batches same-model requests and allows parallel/co-resident models. Both honour the per-request X-LlamaSwap-Priority header."
                         },
                         "settings": {
                             "type": "object",
                             "properties": {
+                                "serial": {
+                                    "type": "object",
+                                    "description": "Tunes how the serial scheduler scores waiting requests: score = X-LlamaSwap-Priority + swap affinity + aging. Ignored unless use is 'serial'.",
+                                    "properties": {
+                                        "agingDivisor": {
+                                            "type": "integer",
+                                            "minimum": 0,
+                                            "default": 60,
+                                            "description": "Seconds a request must wait to gain one priority point. 0 disables aging. Unbounded on purpose: aging is the only term allowed to promote a request across a priority band, which is what prevents starvation."
+                                        },
+                                        "swapAffinityBonus": {
+                                            "type": "integer",
+                                            "minimum": 0,
+                                            "maximum": 99,
+                                            "default": 10,
+                                            "description": "Bonus for a request that can be served without swapping models, so strict priority does not force a cold load on every dispatch. Capped below the 100-point band spacing so it can only break near-ties. 0 disables it, giving strict priority-then-arrival order."
+                                        }
+                                    },
+                                    "additionalProperties": false
+                                },
                                 "fifo": {
                                     "type": "object",
                                     "properties": {
                                         "priority": {
                                             "type": "object",
-                                            "description": "Per-model priority. Keys are model IDs, values are integers (default 0). Higher values are serviced first.",
+                                            "description": "Per-model priority. Keys are model IDs, values are integers (default 0). Higher values are serviced first. Added to the caller's X-LlamaSwap-Priority.",
                                             "additionalProperties": {
                                                 "type": "integer"
                                             }