feat: precompute job registration API and sketch query executor for e2e datapath

[zzylol]

Context

This issue tracks everything ASAPQuery needs to implement so the e2e precomputation data path works.

---

Required API surface

POST /api/v1/precompute/jobs

Register a new precompute job. The controller sends:

{
  "query":       "quantile_over_time(0.99, latency{service=\"web\"}[5m])",
  "granularity": "1m",
  "source":      "backend-collector:4317",
  "sketch_type": "ddsketch",
  "store_path":  "precomputed/latency/p99/5m"
}

Response:

{
  "job_id":     "job-abc123",
  "status":     "created",
  "created_at": "2026-03-26T00:00:00Z"
}

DELETE /api/v1/precompute/jobs/{id}

Deregister and stop a job. Called by the controller on re-plan and rollback. Must return 204 No Content on success.

GET /api/v1/precompute/jobs

List active jobs and their status. Used for operator visibility and health checks.

---

Required internal components

Component	Responsibility
Job registry	Persist active jobs (survive restart); expose GET /jobs for status
Sketch puller	On each granularity tick, fetch accumulated sketch blobs from source (edge collector address)
Query executor	Evaluate query_expr using sketch-native functions (table below)
Result cache	Store scalar/vector result keyed by store_path; serve zero-latency on cache hit
Scheduler	Trigger each job at its granularity interval; honour valid_until TTL on re-plan
Query router	When a user PromQL/SQL query matches a registered store_path, short-circuit to cache instead of DB scan
Metrics endpoint	Expose /metrics (Prometheus) so the controller can observe job health and feed it back into planning decisions

---

Query executor — sketch function mapping

The query_expr strings sent by the controller use these function names:

Function	Sketch type	Semantics
quantile_over_time(φ, metric[window])	DDSketch	Query DDSketch for quantile φ over the window
count_distinct_over_time(metric[window])	HLL	Return HLL cardinality estimate for the window
top_k_over_time(k, metric[window])	CountSketch / CountMinSketch	Return top-k heavy hitters from the sketch

---

Feedback metrics (Prometheus /metrics)

The controller scrapes these to observe actual precompute performance and adjust future planning decisions:

Metric name	Type	Description
asapquery_precompute_job_duration_seconds	Histogram	Wall time per precompute job cycle
asapquery_precompute_cache_hits_total	Counter	Queries served from precomputed cache
asapquery_precompute_cache_misses_total	Counter	Queries that fell through to DB scan
asapquery_precompute_sketch_pull_errors_total	Counter	Failed sketch fetches from source

---

Acceptance criteria

• POST /api/v1/precompute/jobs accepts the request body and returns job_id
• DELETE /api/v1/precompute/jobs/{id} deregisters a job and stops its scheduler
• Registered jobs execute at granularity and cache results under store_path
• quantile_over_time, count_distinct_over_time, top_k_over_time resolve against sketch data from source
• Matching user queries are served from cache (zero DB scan)
• /metrics exposes the four feedback metrics listed above
• Integration test: start ASAPQuery with a controller assigning a job where repeat_every < latency_sla; assert the job is registered, executed, and the result is served from cache on the next matching query