[helly25]

Ordinary comparison summaries now build left and right totals directly, reducing duplicate per-path summary storage. Category summaries and expressions with actions retain their existing behavior.

AG;DR

Based on main after #1056 merged. Synchronization head ae7ec80094471b0d1d94f9aa654d1ffc13741d94 contains main fc2cd316e81d7319aa44912e786e7aad4492cd0a; its complete tree is identical to measured head cc00b95f0842dc8a07f7d936f12badde5f6af85f. Normal synchronization commit/push hooks pass; required CI is running on the current pushed head.

Implementation:

• For overall, type and ext summaries with no category scope, reuse per-root aggregates to construct side-total columns instead of allocating summary vectors/maps under every path and rebuilding the same totals through a category map.
• Reuse the existing independent-filter audit conservatively. Mixed/category scopes, complex summaries and expressions with actions, traversal controls, collections or verification retain the existing per-entry implementation.
• Add no flag-name scans, new effect vocabulary, worker, lock, clock, exception mechanism, dependency or mutation route. Traversal and content observations remain in place.
• This is O16a summary retention only: both complete tree maps remain, and first-output latency is unchanged. Document the still-open O16b completed-frontier interface and O17/O18 concurrency/backpressure boundaries, including empty/ignored completion, ordering, errors, effects, archives and root-relative ignore context.

Evidence and tradeoffs:

• docs/benchmarks/aggregate-comparison-totals-20261102.md and compressed raw data retain source/compiler/build/link/binary identities, all samples, scripts, setup failures and validation logs.
• Per host: 72 native cases x three sessions, broad/deep trees with 10/1,000/10,000 paths, one/three requested workers, discarded warmups, nine rotating rounds/fastest seven. Independent physical count/byte/category oracles and complete baseline/candidate output equality are required.
• Separate in-memory whole-engine population: 30 cases x three sessions x two variants. Confirmation uses an identical baseline participant in eight native cases x five sessions and five in-memory cases x three sessions x three variants.
• Large native totals improve about 16-21% on Linux and 4.6-5.2% on macOS in confirmation. In-memory savings are larger because native startup/traversal/I/O remain. Category/mixed/status controls and small commands retain their measured slower samples; no universal speedup is claimed.
• Peak RSS and unstripped CLI size are reported explicitly. Full tree retention remains linear; this is not a bounded streaming algorithm.

Validation:

• New isolated, read-only VFS tests compare ordinary totals with the existing action fallback across scopes, filters, repeated sinks, directory roots, file/directory changes and empty sides; expected physical counts/bytes are asserted independently.
• 60 complete baseline/candidate formatting comparisons pass per host across plain, Markdown, JSONL, CSV and TSV.
• All three modified translation units pass scoped local clang-tidy. All 323 extension-enabled release tests pass per host, with required Linux FUSE. Five focused ASan suites pass on each host and five Linux TSan suites pass. Current pushed-head CI and normal enforcing hooks remain required before readiness.

Update the canonical performance analysis, TODO, summary design, changelog and release-site mapping. Preserve broader expression, concurrency, archive, metadata, similarity and output work as open rather than claiming it complete from this narrower optimization.