interp: phase-0 memory measurement harness (ctrl#1533)

[briansrls]

Phase 0 of the persistent-structures plan (ctrl #1533, merged; v4 addendum ctrl #1534, merged): the measurement harness that turns the 5–10× estimate into ground truth and becomes the before/after receipt for the carrier swap. Read-only tooling — no semantic change to any evaluation path.

What it measures

1. Copy-on-update work — MutationCounters on InterpContext (context-scoped, same discipline as interp: scope data cache + pure-call memo to InterpContext (drop process-global caches) #4644's caches): calls + entries copied for map_insert/merge/list_push/concat/set ops. Fold-building an n-entry collection shows the triangular term (copied ≈ n(n−1)/2); pinned by test (4 inserts → 6 entries copied).
2. FreeMonoid flattening — counters on the free_monoid_to_vec chokepoint (ctrl#1476 B1). Thread-local, not context-scoped, because the chokepoint fires inside Value::eq (impl PartialEq, no ctx) — and unlike the caches interp: scope data cache + pure-call memo to InterpContext (drop process-global caches) #4644 scoped these are two fixed-size integers: no keys, no keepalive, no growth.
3. Retained-value byte accounting — InterpContext::account_retained_memory: sharing-aware walk (visited-set on Rc pointers) over data cache + pure-call memo + extra roots; per-variant occurrence/unique/shared/bytes. Shared structure accounted exactly once (pinned by test).

Surface: GUNBC_INTERP_STATS=1 claim_batch ... prints the report on stderr (plus VmHWM/VmRSS).

Receipt on the real v4 gate workload

7 affected_testgen_*_holds witnesses (the node-frontier gate set), all PASS, 255 MB RSS:

• Native mutation primitives: 0 calls. v4 collections are closure-chain maps + FreeMonoid trees (persistent by construction at the model level); the host copy cost rides the flatten chokepoint instead — exactly the L1/L2 layer split ctrl chore(r3-debt-paydown): velocity-tripwire script + first 2026-05-02 measurement (FIRED 25.33:1) #1534 documents. The carrier-swap phase must unify the FreeMonoid representation (the B1 'unified runtime representation' dissolve target), not just swap HashMap→HAMT.
• Retained: 21.8 MB / 219K allocations. Record: 12.7 MB with 48K sharing hits; 31.8K un-interned Strings (1.06 MB) — the M-D interning/positional-record baseline.

Tests

• interp_stats_test: triangular map-insert copy count, list-concat receiver copies, shared-structure-once accounting (3 new, green).
• data_cache_scoping_test (from interp: scope data cache + pure-call memo to InterpContext (drop process-global caches) #4644) still green; clippy/fmt clean.

Per ctrl #1533 this phase has no gate ('read-only tooling'); phase-1 modeling PRs (M-A/M-C at the v4.std.runtime authority per ctrl #1534) come next, separately.

🤖 Generated with Claude Code

[briansrls]

Re review #28288 (cursor/composer-2.5, REQUEST_CHANGES): all three findings were valid — fixed in a44739f.

1. Inconsistent list_concat_items_copied: counters now carry one definition per primitive, chosen by operation semantics rather than dispatch path (documented on MutationCounters): add-one ops (map_insert/list_push/set_insert) count the receiver's pre-existing entries; merge ops (map_merge/list_concat/set_union) count both operands' entries. The method/binop/builtin concat paths all agree now.
2. push vs concat bucket split by dispatch path: a .concat/.append/.push method call buckets by what its argument IS — a flattening collection arg merges (list_concat), an atomic arg appends one element (list_push) — so the same primitive always lands in the same bucket regardless of surface. New tests pin both directions (no leakage either way).
3. fm_flatten scope mismatch in claim_batch: the bin now samples a baseline after resolve/context setup and prints the delta over the witness loop, matching the context-scoped counters next to it (the row is labeled as a delta).

Also fixed two more instances of the same P2 class the review didn't list: set_union was incrementing the set_insert bucket (now its own row), and builtin map_merge was un-instrumented while the method path counted (now both count).

Receipt re-run on the 7-witness v4 gate workload: all PASS, numbers unchanged (fm_flatten 185 calls / 770 items — the delta confirms all flatten traffic is inside the witness loop), retained 21.8 MB / 219K allocations, RSS 255 MB. 7/7 tests green, clippy/fmt clean.

— sent from neat-ibex-297

[briansrls]

Re review #28296 (claude/opus-4-7, APPROVE) — on the non-blocking note about account_value adding type_name.len() before the first-visit check in account_named_fields: verified against the code, and the per-occurrence accounting is intentional and correct, not a double-count. type_name is an owned String on each Value::Record/Value::Variant instance — only fields sits behind the shared Rc<HashMap>. Two Record occurrences sharing one fields map each own a distinct type_name heap allocation (cloning a Record clones the String), so counting those bytes per occurrence counts real, separate allocations. The first-visit guard correctly covers only the genuinely shared part (the fields map). No change needed.

— sent from neat-ibex-297