Repository source: docs/plans/qdb_build_time/qdb_build_time_step4_implementation_summary.md
State: Eight approved stages locally complete on feature/qdb-build-time; production replay and maintainer acceptance pending. Planning commit: f526da98. Frozen baseline: 4abda3db, /tmp/qdb-build-time-baseline. Inherited state: Prior pagination cycle complete; preserve unrelated audit drafts, ledger edits and prior-cycle host report. No merge, push or production changes in this cycle yet. Study: railspike-effort, R1/P2, v3/qdb_build_time; original v1/A 0.50 retained, selected v2/C 1.25. Single exposed Codex rater; reliable measured labor unavailable.
Entry cycle estimate: 1.25 Revised cycle estimate: 1.35 (S7, same scope) Retrospective cycle assessment: 1.40
Stage 1 — Reproducible baselines and parity fixtures
Original stage-work estimate: 0.15 Retrospective stage-work assessment: 0.18
Changes: Fixed corpus generator and separate-operation harness; frozen full row outputs beside baseline checkout. Added production rendering batch timings.
Verification: Frozen baseline activity suite 8/8, normal CLI static build 1/1; instrumented CLI static test 1/1. Edge and 10000-event full row streams repeat exactly; empty/missing Git produce empty streams. Valid 5215-profile baseline: 115.4633 seconds, 10 MiB PHP peak, 5215 files. PHP lint and diff whitespace checks pass. Failed exploratory samples and fixes retained in findings. Stage 7 owns repeated matched acceptance.
Notes: Originals for subsequent stages: S2 0.16, S3 0.15, S4 0.20, S5 0.23, S6 0.16, S7 0.15, S8 0.10. Scope unchanged from approved Step 3; no revisions.
— Codex · 2026-10-30 UTC
Stage 2 — Explicit profile delivery control
Original stage-work estimate: 0.16 Retrospective stage-work assessment: 0.13
Changes: Added site-mutable FORUM_STATIC_PROFILE_PAGES_ENABLED, default true, adjacent to static detail control. Only full-build profile enumeration is conditional. Registered nullable private schema metadata, excluded from template defaults like other site flags so templates cannot override canonical settings.
Verification: Default/on profile HTML byte equality; off produces zero profile artifacts while thread flag remains independent; all profile rows and offline snapshot retained. Configuration creation/refresh and schema tests passed. Updated registry inventory after expected first-run failure; evaluator/presentation suite 23/23 passes. Earlier configuration/control run: other 32 checks passed.
Notes: Runtime/stale-artifact transitions and browser flow belong to Stage 3. Dependencies added: zero. Scope/estimate unchanged; labor unmeasured.
— Codex · 2026-10-30 UTC
Stage 3 — Profile transitions and recovery
Original stage-work estimate: 0.15 Retrospective stage-work assessment: 0.15
Changes: Disabled profiles bypass both public-root and release artifacts in PHP; shared refresh excludes copied profiles; single-route regeneration skips them. Added operator mode/rebuild/recovery guidance and concept entry.
Verification: Four focused PHP checks pass: both transitions, stale public/current artifacts, repeated runtime reads, changed profile, failed candidate retaining prior release and retry, shared refresh, approved and pending viewer access. Real Chromium browser flow passes directory → account → profile, disabled runtime delivery and flag presentation with JavaScript on/off, phone/desktop and 16 profile themes; no page overflow. Initial browser failures were incorrect test selectors (directory links to accounts first; flag title contains badges), corrected without product changes. Enabled artifacts remain covered by Stage 2 and transition tests. Diff/lint checks pass.
Notes: Direct physical HTML requests bypass PHP, documented for old exported public trees. No dependency delta, no UI navigation additions. Original scope/estimate unchanged; labor unmeasured.
— Codex · 2026-10-30 UTC
Stage 4 — Cheaper streamed commit indexing
Original stage-work estimate: 0.20 Retrospective stage-work assessment: 0.18
Changes: CommitHistoryReader streams raw tree changes from a file-backed process spool, counts files without line diffs, and yields one metadata row at a time. Git must finish successfully before rows are inserted; failed/empty/missing history preserves the former empty result. Existing SQL and schema unchanged.
Verification: Exact complete rows match frozen edge and 1001-commit outputs. Two new tests pass across rename settings true/false/copies and four merge settings, including mode-only, empty-file, binary, Unicode/tab/newline names, corrupt objects and absent/empty Git. Initial integrated ordinary measurement: 0.023859s, 2 MiB PHP peak (acceptance repeats remain Stage 7). Three timing checks pass; testProgressSeparatesParsingDerivationAndWritingWithBoundedCounts fails identically in untouched baseline and branch and is recorded as pre-existing, not a pass.
Notes: Temporary spool prevents partial output on Git failure and retains heap bounds; no rename/merge-policy overrides. Dependency delta zero; original scope/estimate unchanged; labor unmeasured.
— Codex · 2026-10-30 UTC
Stage 5 — Bounded manifest reuse in evidence refresh
Original stage-work estimate: 0.23 Retrospective stage-work assessment: 0.22
Changes: Operation-local file-backed SQLite lookup streams batched Git manifests, retains current small commit group on disk, and fetches only source/signature/key paths into PHP. Shared commits survive adjacent 200-row batches. Transactions prevent partial manifests; failed batches retain per-commit fallback. Full commit-detail requests keep their existing path.
Verification: All eight ActivityProjection tests pass. Frozen 10000-event/20000-path output is byte-identical; preliminary optimized run 0.579404s, 4 MiB PHP peak versus baseline 16 MiB. New tests pass exact 501-row evidence across three batches with one traced manifest process, small-commit batching, explicit/destructor cleanup and an injected partial-output/nonzero Git failure with no partial evidence or repeated per-event retry.
Notes: No core read-model queries/schema changed; scratch is private and not exported. Temporary-store cache budget is 1 MiB; final matched acceptance remains Stage 7. Dependencies zero; original scope/estimate unchanged; labor unmeasured.
— Codex · 2026-10-30 UTC
Stage 6 — History maintenance and failure parity
Original stage-work estimate: 0.16 Retrospective stage-work assessment: 0.14
Changes: Added complete signed-event/manifest equivalence checks (record, signature and key), rename/new-operation isolation and injected scratch-write failure/rollback/cleanup coverage. Updated rebuild concept with operation lifecycle and failure behavior.
Verification: Focused maintenance/reader/candidate suite 48/48 passes: full and incremental refresh, key arrival/removal, missing/broken history, bootstrap fallback, standard and Forte activity/RSS, current/blob/commit source routes, normal rebuild CLI, candidate promotion failure/rollback/retry, provenance batching, exact signed evidence and scratch cleanup. Existing schema recovery tests remain green.
Notes: Core table definitions and queries are unchanged; only private scratch has a new schema, so deployed schema stays 39. No dependency delta, trust decision or export-format change. Original scope/estimate unchanged; labor unmeasured.
— Codex · 2026-10-30 UTC
Stage 7 — Separate and combined performance acceptance
Original stage-work estimate: 0.15 Revised stage-work estimate: 0.22 Retrospective stage-work assessment: 0.27
Changes: Serial matched fresh-process baseline/isolated/combined harness, retaining CPU/RSS/phase/artifact/hash/process records. Initial many-small-commit evidence samples exposed scratch fsync and per-query locking overhead (0.56s baseline versus 3.74–4.99s new first samples), despite identical outputs. Correcting within the approved bounded-scratch target: one operation transaction with per-load savepoints, disposable-store synchronous-off/exclusive locking, and at most 512 requested-path results including misses. Journal rollback remains enabled; no durable data uses these settings.
Verification: Initial failed/regressing samples remain in /tmp/qdb-build-time-results/matched/samples.txt and are not acceptance. Final scratch variant must repeat large/small evidence comparisons and failure tests. Profile/commit/runtime samples do not execute scratch and remain independently usable.
Estimate revision: 2026-10-30 UTC, same v2/C scope and single exposed Codex rater, R1/P2; revised entire S7 0.22 due to observed scratch overhead and repeat measurement work, not scope growth. Independent cycle projection revised from 1.25 to 1.35; originals retained, labor unmeasured. No outcome claimed before gates.
— Codex · 2026-10-30 UTC
Stage 7 completion: Three fixed target gates pass with exact output equality: ordinary commit indexing 1.0362 → 0.0219s combined median; large shared-commit evidence 1.4645 → 0.1970s; profile enumeration/rendering 69.0015 → 0.0024s with zero profile files. Prepared-model full publication 85.6020 → 15.9964s and normal full CLI 1.74 → 0.41s, with exact posts/profiles/activity/commits parity and activation. Runtime profile budget passes (0.0298s versus 0.0565s allowed), while static hits remain faster (0.0093s). PHP peaks never exceed matched baseline; all runs capped at 128 MiB, maximum externally measured RSS 53412 KiB. The supplemental 1000-small-commit evidence case costs 0.0407s (8.7%) more and has no demonstrated improvement; this limitation is retained rather than generalized away. Final tuned lookup/projection/transition/all-site runtime-asset suite 14/14 passes. See measurements and all retained samples. No production replay or full-suite claim.
— Codex · 2026-10-30 UTC
Stage 8 — Documentation and host replay handoff
Original stage-work estimate: 0.10 Retrospective stage-work assessment: 0.12
Changes: Final findings, feature-cost catalog, profile-delivery and history-evidence concepts, operator toggle/rebuild/recovery/host replay guidance, plan navigation and scoped study ledger. Stage commits: planning f526da98; S1 cded5a7f; S2 e145fbfa; S3 bed764ef; S4 8f4f4905; S5 42e9867e; S6 38c41ea6; S7 5d78849d; S8 is this documentation closeout commit (exact hash recorded in final debug handoff).
Verification: Final focused PHP suite 80/80; previously recorded real Chromium profile/control flow passes JS on/off, phone/desktop and 16 profile themes. All touched concept/plan local link targets resolve; PHP/Python syntax and diff whitespace checks pass. All three frozen R1 hashes reverified. Dependency inventory delta zero; schema 39 retained. No production deployment, full-suite or physical-device claim.
Notes: Original scope/estimate unchanged; reliable labor unavailable. User's original build argument list was not supplied, so host guidance preserves their command/settings rather than inventing it. QDB PHP 8.3 replay compares recorded 314.961s total, 39.1605s commits and 44.503s evidence, with corpus/configuration differences recorded.
Closeout
Local implementation and primary timing/memory/runtime gates pass; normal CLI build/activation and data parity are verified. Default profile prebuilding remains enabled for every site; PHP-hosted operators can disable Static profile pages or set FORUM_STATIC_PROFILE_PAGES_ENABLED=false consistently for builder and frontend. Static-only hosting retains enabled mode. No canonical records, trust requirements or existing profile routes were removed.
Known limitations: Supplemental small-commit evidence has a measured 0.0407s/8.7% overhead, and runtime profiles remain slower than static hits. ReadModelBuilderTimingTest::testProgressSeparatesParsingDerivationAndWritingWithBoundedCounts fails identically in the frozen baseline and branch; it is outside the passing 80-test focused final suite. Production timing and maintainer comfortable-fit/acceptance are pending. No merge or push performed in this cycle.
Study: Same v3/qdb_build_time v2/C, baseline 4abda3db, R1/P2, single exposed Codex rater. Original v1/A 0.50 and v2/C entry 1.25 preserved; prospective cycle revision 1.35 and S7 0.22 recorded at discovery. Retrospective whole-cycle 1.40 is independent of non-additive stage assessments. Actual labor and independent rating unavailable; no invented maintainer judgment or accepted-outcome increment. Unrelated audit drafts, their ledger entries and the previous cycle's host-report edit remain uncommitted and preserved.
— Codex · 2026-10-30 UTC