memory(gc): copied minor subtracts from-space high-water from an exact live census

[proggeramlug]

Summary

The new exact live-byte census can undercount after collector modes alternate. A full/non-moving sweep publishes exact aggregate live bytes, excluding dead objects left as holes in partially-live Eden blocks. The next copied minor derives its census by subtracting copying_from_space_in_use_bytes(), which is the sum of block bump offsets and therefore includes those already-excluded dead holes.

That subtracts the same dead space a second time. saturating_sub can hide the arithmetic failure and, when the young high-water exceeds unrelated old live bytes, erase old-generation occupancy from heapUsed and major-GC pacing altogether.

Concrete accounting

After a non-moving sweep, suppose:

• old/long-lived live bytes = 40 MiB
• Eden high-water = 10 MiB
• actual live Eden objects = 1 MiB (9 MiB dead holes in partially-live blocks)
• published exact census = 41 MiB

On the next copied minor, if the 1 MiB survives, current code computes:

41 MiB - 10 MiB from-space high-water + 1 MiB survivors = 32 MiB

The correct result is still 41 MiB. With less than 9 MiB of unrelated live data, saturating subtraction reduces the non-from-space component to zero.

Evidence

• copying_from_space_in_use_bytes() sums Eden and active-survivor block offsets; it is a high-water/placement quantity, not live occupancy.
• The copied minor captures that high-water and the previously exact live census.
• It subtracts the full high-water, then adds copied/promoted survivors.
• Non-copying collection publishes sweep.arena_live_bytes, the object-walk census.
• arena_live_allocated_bytes() explicitly describes the stored value as an exact post-GC census.
• Current tests cover copied minors after bump allocation/promotions, but do not create a partially-live from-space high-water via a preceding non-moving/full sweep.

Impact

• process.memoryUsage().heapUsed and V8-compatible heap telemetry can underreport live memory.
• Major/full-GC pacing can be delayed because it consumes the same undercounted metric.
• The error persists in the next census and can grow across mode transitions until a later whole-heap object census repairs it.
• The use of saturating arithmetic makes the invariant violation silent.

Acceptance criteria

• Track or derive live from-space bytes, distinct from block high-water, when converting an exact aggregate census through a copied minor.
• Preserve exact non-from-space live bytes across copied minors, including prior promotions, long-lived objects, old holes, and partially-live Eden blocks.
• Add a regression sequence: seed an old live cohort; run a non-moving/full sweep leaving one young survivor in a partially dead block; run a copied minor; assert heapUsed == old_live + copied/promoted_live byte-for-byte.
• Add an internal checked/debug invariant that prevents a high-water subtraction from silently consuming unrelated generations.
• Keep high-water/reserved telemetry separate for fragmentation analysis.

Audit context

Found during the v0.5.1481 re-audit of #7879/#7886. The original high-water-as-live defect is fixed for direct post-collection reads; this issue is a cross-mode derivation error in the replacement census.