fix(memory): retain cgroup pressure after allocator trim

[ScriptedAlchemy]

Summary

Supersedes #2596 (agent/master-unpushed-20260928). That pull request is left open.

• After an allocator trim, admission keeps the cgroup-aware sample. Publishing the post-trim unreclaimable RSS was clearing a kill-line observation.
• A failed checkpoint read returns the standing pressure verdict. Callers were treating None as room to keep allocating.
• The parent-carried lexical artifact path is withdrawn. It sealed incomplete source receipts under the current format revision. A carried marker is incompatible and the next build is a full rebuild.

Motivation

#2596 conflicted with current master and its six Linux Hauler lanes failed against its pinned base. Diffed against current master first:

• fix(memory): allocate SQLite and tree-sitter from mimalloc #2773 routes SQLite and tree-sitter through mimalloc and deletes the periodic glibc trim. The pressure reclaimer still publishes sampled_process_resident_bytes_v1() (unreclaimable RSS) after a trim, so cgroup pressure is still dropped.
• refactor(memory): size admissions from one resident-memory view #2785 sizes admission from one headroom view. That test is kept.
• fix(memory): charge resident owner pages, admit headroom #2790 charges resident owner-heap pages. It does not touch the trim reclaimer.
• The carried lexical path from fix(code-index): resume lexical text from the parent artifact #2437 is still on master. fix: fit Windows paths, CRLF field sites, and change witnesses #2540 only adjusted Windows paths in it.

Changes

• register_process_allocator_pressure_reclaimer_v1 re-reads through sample_for_checkpoint after the trim instead of publish_post_reclaim_observed_resident_bytes.
• sample_for_checkpoint returns ResidentMemoryPressureStateV1. A failed sample leaves the last published verdict in place. Checkpoint and graph-publication callers use that verdict.
• Deleted the parent-carry delta (artifact/delta.rs), the source restriction that fed it, and restart_after_sigterm_rebuilds_only_changed_files. That test asserted the withdrawn optimization (only the edited file is rebuilt). wait_for_refresh_in_flight existed only for it. Unchanged files stay searchable because the successor build is a full text artifact. Existing carried staging files fail layout verification and are rebuilt.

Test plan

• cargo fmt --all -- --check — exit 0
• cargo clippy -p tracedecay-runtime-core -p tracedecay-application -p tracedecay-store-runtime -p tracedecay-code-index -p tracedecay-query -p tracedecay-code-index-runtime --all-targets -- -D warnings — exit 0
• cargo clippy -p tracedecay --test daemon_suite --features test-helpers -- -D warnings — exit 0
• cargo test -p tracedecay-runtime-core --lib resident_memory — 51 passed
• cargo test -p tracedecay-application --lib code_index:: — 2 passed (checkpoint_sampling_failure_preserves_known_pressure, a_checkpoint_stops_at_the_watermark_without_reclaiming)
• cargo test -p tracedecay-store-runtime --lib resident_memory_trip_is_permanent_for_one_publication_attempt — 1 passed
• cargo test -p tracedecay-query --lib — 275 passed, 1 ignored
• cargo test -p tracedecay-code-index --lib — 270 passed
• cargo test -p tracedecay-code-index-runtime --lib — 562 passed, 1 ignored
• cargo test -p tracedecay-code-index-runtime --lib staging_sidecar_tests — 1 passed
• cargo test -p tracedecay-code-index-retention --lib staging_sidecars_share_their_staging_artifact_liveness — 1 passed
• cargo test -p tracedecay-code-index --lib lexical_page — 12 passed
• cargo nextest run --workspace --no-fail-fast not run
• Changelog left to release-please
• No secrets included
• No breaking external API; sample_for_checkpoint is an in-workspace signature change and every caller is updated

bounded_restore_refuses_a_same_inode_rewrite_witness failed once inside the first full tracedecay-query lib run and did not reproduce across eight consecutive reruns of that suite. The test and the restore witness path are unchanged from master; the assertion is still Corrupt.

  

[changeset-bot]

⚠️ No Changeset found

Latest commit: 2d27d53

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR