feat(cost): estimate disk and object-store operation counts

[zzylol]

Why

Equal scan bytes can require different storage request counts and costs. Addresses #345 independently of #349 and #346.

What and how

Estimate disk reads/writes and object GETs/PUTs as sum(ceil(extent_bytes / bytes_per_request)) × executions. Count shared physical nodes once. Require fresh, complete node/statistics evidence; scan read extents must match authoritative source-read bytes. Add separately calibrated request costs to ranking and export totals, per-node terms, and provenance.

Before this PR

A shared scan of two 40-byte objects over three evaluations reports 240 scan bytes but no GET estimate.

After this PR

With 32-byte requests it reports 12 GETs; a build-once scan reports four. A 160-byte write with 64-byte requests adds three writes per execution. JSON-to-ranking/export coverage verifies 160 raw GETs versus 10 candidate GETs, then explicit writes that reverse the preference. A request-only objective works with zero base CPU/byte coefficients.

Verification

• Independent review completed; added request-only objective support with a failing-before/passing-after regression.
• Passed cargo test --workspace --locked, including current main's test(promql): verify TopK planning and numeric reference results #351 tests.
• Passed cargo clippy --workspace --all-targets --all-features --locked -- -D warnings, formatting, and diff checks.
• Tests cover per-extent rounding, monotonicity, integer overflow, shared reads, explicit writes, retained-state multiplicity, stale/missing/rebound evidence, strict JSON, ranking/export, and rejection of absent/invalid/all-zero objectives.

The examples are executable analytical checks, not measured I/O. Screenshots and empirical runtime benchmarks: not applicable; none were produced.

Architectural decisions

An optional deployment-snapshot profile binds request sizing and storage accesses to complete physical node/statistics records. Counts remain separate from bytes until calibration. Missing profiles preserve the existing objective without fabricating request counts.

Limitations and follow-up

The payload model excludes metadata, retries, speculative prefetch, multipart control requests, and empty-object creation. Lifecycle-specific maintenance and cache integration remain separate. Rust snapshot constructors gain an optional profile field; existing JSON remains supported. This branch includes current main and is not stacked on the other cost PRs.

Human review — do not complete with an agent

• The MVP boundary is correct.
• New conceptual layers or public interfaces are necessary.
• The before/after description matches the intended product behavior.
• Human reviewer:
• Decision and rationale:

[zzylol]

@Selvomega

[zzylol]

Review and fixes pushed in c3c5846 (plus main merge d273191, including #357 shared resource compatibility). Finding P2: an empty/whitespace base calibration version was accepted, including storage-only objectives. The constructor now rejects missing provenance before snapshot lookup; the new regression was observed failing before the fix and passing afterwards. Additional passing tests cover zero extents/all four request kinds, multiplicity and cross-node overflow, and node/source/statistics rebinding. Those cases confirmed the original storage arithmetic rather than identifying additional bugs. Validation passed: cargo test --workspace --locked; cargo clippy --workspace --all-targets --all-features --locked -- -D warnings; format/diff checks (363 mapping unit tests and 7 storage integration tests included). Normal push only; no PR merge or human-review attestation.

[zzylol]

Updated in e44e0f2: truly rebased onto main c3410d1 with linear PR history and all previous correctness fixes preserved. StorageResources now has one canonical definition in asap_types::resources, covering disk reads/writes and object-store GET/PUT counts; the mapping module re-exports the shared type for import compatibility. Estimation, evidence binding, and calibration remain in mapping. Added serialization, per-dimension checked-add overflow, shared-type identity, and missing-profile regressions. Full workspace tests --locked and strict Clippy (workspace, all targets/features) pass, including 364 mapping unit tests, 8 storage integration tests, and 165 types tests. Updated with an explicit force-with-lease; not merged.

[zzylol]

Follow-up: split the remaining shared definitions into resources/cpu.rs (ModeledCpu and MeasuredCpu), resources/measurement.rs (Measurement), and resources/physical.rs (PhysicalResources and MeasuredResources). resources.rs retains module declarations, re-exports, and cross-resource contract tests. Existing root import paths, fields, units, defaults, and JSON formats are unchanged. Full workspace tests --locked and strict workspace Clippy across all targets/features pass.

[zzylol]

Rebased onto merged #353 (main 53e76d4), retaining cache and storage evidence fields and both sets of exported cost annotations. A new regression exposed an unsupported combination: aggregate cache hit ratios cannot determine independently rounded per-extent storage requests. Comparison now fails closed for CacheProfile::Evidence plus storage_io until cache-aware extent evidence exists; explicit NoCache plus storage remains supported. The regression failed before the guard and passes afterward. Full local validation and CI are being run before merge.