Repository navigation
Entry-view assembly memos: measured no-go (cohort receipts only) - #7867
Merged
Merged
Conversation
Wire closure/root assembly env and rewire memos through reconcile, fix im::Vector type mismatches, and drop the per-module census RefCell path that panicked at runtime. Co-authored-by: Cursor <cursoragent@cursor.com>
Delete the dead type_name_export_index_memo path, fold variant-eligibility and source_indices derivation into memo digests with documented notes, and enroll discriminating witness controls in entry_view_assembly_memo_witness_test. Co-authored-by: Cursor <cursoragent@cursor.com>
Representative 50-entry cohort: 1 cross-grain hit in 106 builds — closure/root coarse memos retain the win; the v2 digest recomputed eligibility on every miss. Remove the stale claim_batch receipt hook that broke fmt at 257175f. Co-authored-by: Cursor <cursoragent@cursor.com>
Delete the hollow fallback counter (no fallback arm exists), add memo hit/miss and key-count instrumentation for retention receipts, document variant-base cache key completeness at closure/root grain, and enroll the discriminating witness. Co-authored-by: Cursor <cursoragent@cursor.com>
Name why bare source_tree_root suffices for tree_bare_census_for_root within a fixed MultiEntryIndex, and when incremental index population re-opens the key. Co-authored-by: Cursor <cursoragent@cursor.com>
Fixed post-#7836 50-entry cohort, interleaved base/after arms on one executor: −1.5% wall, −4.7% assembly, +1.8% peak RSS; closure_env and rewired memos 0 hits on this roster. Raw arms and derived summary.json retained under docs/plans/receipts/entry-view-assembly-direction/. Co-authored-by: Cursor <cursoragent@cursor.com>
Cohort A/B on the post-#7836 roster: −1.5% wall, −4.7% assembly, +1.8% peak RSS; closure_env and rewired memos 0 hits / 50 misses. Delete process-lifetime memos, claim_batch receipt line, and witness controls — retain measurement receipts only. Co-authored-by: Cursor <cursoragent@cursor.com>
gunbai-bot
Bot
changed the title
The new receipt md was unreachable from registered roots, so doc_graph_has_no_orphan_docs failed on PR #7867. Co-authored-by: Cursor <cursoragent@cursor.com>
Make the load-bearing finding — a working cache that still does not move wall — the headline; state cohort bounds and equivalence honestly before undraft. Co-authored-by: Cursor <cursoragent@cursor.com>
Raw terminal colour codes made the PR diff undiffable for review tooling (resolve-pr-diff + gh api fallback both failed). Co-authored-by: Cursor <cursoragent@cursor.com>
Receipt cited an empty binary hash file; populate claim_batch digests from summary.json. Floor-prep follow-on now leads with root_env 92% hit / −1.5% wall per receipt headline. Co-authored-by: Cursor <cursoragent@cursor.com>
- Fix derive_summary.py to retain closure_env/root_env/rewired breakdown (was flattening to last-wins rewired-only dict). - Commit per-arm TSV + stderr artifacts extracted from remote_cohort_run3.log (fill-composition precedent); regenerate summary.json. - Add SCAFFOLD + dissolve-on headers to run_all.sh and run_ab_cohort.sh. Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
Author
|
Addressed review 49494 (cursor/composer-2.5 REQUEST_CHANGES): derive_summary.py / summary.json — Confirmed: Shell scripts — Added Per-arm artifacts — Committed Head: pushing now; CI will re-run. — sent from calm-crab-696 |
gunbai-bot Bot
pushed a commit
that referenced
this pull request
Aug 7, 2026
Conflicts and how each was resolved: docs/plans/floor-prep-tax-program.md — took main's side. Its P1 paragraph contains this branch's text verbatim plus a newer verdict (entry-view assembly memos REJECT rather than broader shared-entry-view program OPEN) and an added P1 follow-on recording PR #7867's rejection. This branch's copy was the stale half. src/v1/stage0/src/v1_compiler_infer.rs — NOT hand-resolved. This is a generated artifact and the merge emitted `.githooks/generated-artifact-merge: No such file or directory`, so the driver was absent in this clone and git text-merged a file that must never be text-merged. The .dag authority (src/v1/04_infer.dag) merged cleanly, so the seed was reset to main's compilable version and regenerated from the merged source: regen_stage0 wrote 134 generated stage0 files, verified by tree diff rather than by the tool's success message. The merge also brings in the missing generated-artifact merge driver itself. Verified: no conflict markers remain in the tree; the regenerated seed compiles (cargo build -p v1-compiler --bin claim_batch, exit 0). Integrated with a merge commit rather than a rebase, per the repo's squash-merge policy — identical end state, no force-push. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
PR B proposed three process-lifetime memos on
MultiEntryIndexto reuse entry-view assembly across witnesses in one floor process (closure_assembly_env_memo,root_assembly_env_memo,rewired_modules_memo). Fill-composition overlay (#7836) is unchanged onmain.Measured verdict: no-go. The mechanism is deleted from this branch; what remains is the cohort receipt and write-up documenting why.
Cohort A/B receipt (post-#7836 50-entry roster)
One tree, one variable (
base-arm-revert.patch), interleavedbase-r1 / after-r1 / base-r2 / after-r2on one BuildBuddy executor. Evidence:docs/plans/receipts/entry-view-assembly-direction/·summary.json· receipt write-up.Headline:
root_envhit 92% (672/57) and wall still moved only −1.5% — entry-assembly memoization as a class is not where time goes, even when the cache works.closure_env/rewired0 hits are secondary (keys do not repeat on this roster).Outcomes: 48 PASS / 2 FAIL, identical failure names across all four arms (pre-existing hermetic refusals).
Benefit bar: ≥10% or ≥3 minutes floor recovery; peak ≤ post-#7836 envelope (~5.8 GiB). Miss on all performance axes.
Representative affected-floor run was not started — cohort miss is decisive for this direction.
Disposition
origin/mainintegrated post-measurement.[entry-view-assembly]instrumentation, andentry_view_assembly_memo_witness_testremoved.mainis receipt artifacts only.Receipt hygiene note
remote_cohort_run3.logwas captured with terminal colour codes; 478 ANSI escapes silently broke PR review tooling (resolve-pr-diff+gh apifallback). Stripped to plain text before merge. If you commit a captured execution log, strip escapes first (sed 's/\x1b\[[0-9;]*m//g'). Candidate for a future corpus check.Test plan
run_all.shwithCOHORT_ONLY=1, remote executor)main