fix(store): make trackedBytes accurately enforce maxMemoryMB #113

Upstream source

Evaluate and adapt Kpa-clawbot/CoreScope#2035: correct the in-memory store accounting used by packetStore.maxMemoryMB.

Current fork evidence

Verified against fork master 1bdf088e:

• StoreTx has no persistent record of how many bytes it was charged.
• A transmission is charged before pickBestObservation assigns PathJSON, so path-dependent byPathHop and O(path²) subpath costs are omitted.
• Eviction later recomputes the estimate after the path exists and can subtract more than was originally added, drifting trackedBytes downward.
• The estimate omits the decoded cache, separately allocated dedup-key strings, and relevant observation strings.
• docs/go-migration.md still incorrectly says the Go server ignores packetStore.maxMemoryMB.

An underestimated counter can prevent intended eviction and leave the real heap pressed against its runtime/container limit.

Fork-specific adaptation

Do not cherry-pick blindly. The fork pins unsafe.Sizeof(StoreTx{}) to 320 bytes for route-mask padding and has additional indexes/caches. Preserve that layout with a justified design or update the contract deliberately, and recalibrate constants on this branch.

Acceptance criteria

• Record, directly or equivalently, the exact estimate charged per transmission.
• Apply only deltas when selected paths or other charged allocations change.
• Eviction subtracts exactly the amount charged.
• Include path/subpath indexes, decoded-cache allowance, dedup-key allocations, and relevant observation strings.
• Deterministic insert → path assignment → recharge → eviction returns the counter to exactly zero.
• Mixed insert/update/eviction tests show no drift; memory eviction reliably crosses high and reaches low watermarks.
• Cold-load budgeting and documentation use the revised estimate.
• Existing route-mask, index, and cache behavior remains intact.

Verification

Run focused and full server tests under -race, static checks, before/after tracked-vs-heap measurements on a read-only production-sized copy, cold-load/eviction measurements at representative limits, and insert/eviction performance checks.

No deploy or live-environment mutation is part of this issue.