Full mark-sweep marks the remembered set as roots, retaining young objects through DEAD old-gen pages #9629

What

A full mark-sweep runs RootScanSubphase::RememberedSet, which marks every dirty old-gen page's slots as roots. In a full trace the old generation is already traced from real roots, so the only thing this adds is retention: it marks through old objects that are themselves dead.

Evidence

Fixture (compiled, Linux x86_64): build 10,000 strings into an array large enough to be promoted, drop the array, call gc().

PERRY_GC_TRACE=1 for that collection:

remembered_set {"dirty_pages_scanned": 20, "dirty_slots_scanned": 10000,
                "newly_marked": 10000, "old_objects_considered": 1, "valid_roots": 1}

newly_marked: 10000 is the entire dropped population, and old_objects_considered: 1 is the dead backing array. A heap census taken inside the same cycle confirms the 10,000 strings are classified live, not dead, with the remembered-set count accounting for all of them.

The dead array survives because of #9628, so its pages never leave the dirty set, and every later full collection re-marks the same 10,000 strings. The two interact: #9628 keeps the owner alive, this keeps everything the owner points at alive.

Why it looks redundant for a full cycle

mark_remembered_set_roots validates that the dirty page's object is a plausible pointer (valid_ptrs.contains), not that it is live. That is the right check for a minor, where the old generation is deliberately not traced and old→young edges genuinely are roots. A full trace visits the old generation from the real root set, so a live old object's young children are reached anyway, and the only objects this phase adds are those reachable exclusively from dead old objects.

Impact

Any old-gen object that dies while holding young references keeps them alive until a collection that does not consult the remembered set, which currently is no collection at all. On the compiled claude-code TUI this is part of why an idle process holds a large dead set that only allocation pressure clears.

Not proposing the fix here

Gating the subphase on self.minor.is_some() is the obvious shape, but the remembered set is also the input to evacuation and to the rebuild that runs in AtomicFinalize, so whether a full cycle can simply skip it is a question for whoever owns the generational invariants. I have not tested that change.

Found while validating a heap-census instrument (PERRY_GC_CENSUS, branch feat/gc-census).