Feature plan: Step 1 · Step 2 · Step 3 · Step 4

Debug mode: enabled Effort study: railspike-effort / v3/activity_load_performance; ledger Extended estimates: enabled Cycle estimate: 0.95

Entry: 2026-10-28; Step 3 approved, branch feat/activity-load-performance, planning commit 2028f6d3a. Scope v3/C, baseline 7ef735dd21b3329f18716bf3d5484baab2b2b71b, R1/P2, Codex single exposed rater. Entry total 0.95 reaffirmed. Original S1–S7 stage work 0.08/0.18/0.18/0.15/0.10/0.12/0.15 copied from approved plan (not summed). No new dependencies; production deployment/acceptance remain separate.

Stage 1 - Baseline and read-cost contract

Original estimate: 0.08 Retrospective stage work: 0.08

• Changes: Added repeatable read benchmark using the existing disposable bulk-commit fixture; preserved investigation/writer audit. It measures data-layer work, not HTTP/paint or production.

• Verification: Fixture generation and 20 samples per standard/Forte/detail path passed; baseline timings recorded below. Fixture includes 205 long-path events sharing 2,300 unrelated files; existing signed/missing-history tests supply complementary cases.

• Notes: Existing production observations remain the user-facing baseline; local fast reads do not reproduce server contention. No live runtime/database mutated. Stage 1 owns shared benchmark setup; later stages verify correctness and request cost after restart.

— Codex · 2026-10-28 UTC

Baseline data-layer p95: standard 7.457 ms, Forte 5.018 ms, detail 4.921 ms; 851,968 bytes / 210 events. Warm manifest cache and unsigned bulk data make this a structural comparison, not a production latency prediction.

Stage 2 - Compact rebuild evidence

Original estimate: 0.18 Retrospective stage work: 0.18

• Changes: Schema 32 adds compact evidence, pending index and source-field invalidation. Transactional rebuild computes relevant paths once per commit group and reuses stored signature statuses. Links/HTML/full manifests are not stored.

• Verification: ActivityProjectionTest + schema/prior-index startup checks 6/6; earlier standard pagination checks 11/11. Actual ./v3 rebuild on disposable bulk fixture passed. Prior schema test removes the new column/index/triggers and verifies normal Activity startup rebuilds before reading it.

• Notes: Initial schema inventory assertion failed as expected and was updated; rerun passed. Live readers still use their prior paths until stages 5/6. Stage 1 commit 3e924f927. No dependency changes.

— Codex · 2026-10-28 UTC

Stage 3 - Evidence follows writes

Original estimate: 0.18 Retrospective stage work: 0.18

• Changes: Incremental metadata completion now refreshes pending evidence inside the write transaction, including direct flag inserts. Source-less handoffs receive empty evidence through the schema trigger. Identity/key arrival invalidates only that identity's prior evidence and refreshes its stored reaction status; imports/deletions use the already-wired rebuild path.

• Verification: Projection/reaction/label/candidate recovery tests 13/13; new test checks incremental/rebuild parity and rollback, plus direct source-less insertion.

• Discovery/fix: Parity exposed an existing reaction updater defect: refreshing a target assigned the newest commit to all older reaction events. It now resolves each source's actual commit, matching rebuild and preserving correct evidence. This is required evidence correctness within scope, not a new feature.

• Notes: Stage 2 commit d372c401a. No reader switch yet. Full signed/identity/retraction acceptance continues in Stage 7.

— Codex · 2026-10-28 UTC

Stage 4 - Transactional counts and positions

Original estimate: 0.15 Retrospective stage work: 0.15

• Changes: Schema 33 adds maintained totals with/without profile-name events, commit totals, a source-less cursor revision, and standard/Forte ordering indexes. Triggers cover all Activity writers and post visibility changes; standard pagination no longer reads/hashes every source-less event. Count read switch remains Stage 6.

• Verification: Projection/schema checks 7/7; standard navigation suite 11/11. Membership parity covers malformed/scalar/case-varied tags, feature-name variants, direct inserts/updates/deletes and post hide/retract/delete. Prior schema 32 startup rebuilds missing counters; actual CLI rebuild on the bulk fixture passed.

• Notes: Stage 3 commit 307cb5633. Trigger maintenance is within the same transaction as mutations; no new worker or configuration. Schema inventory assertion updated for new objects.

— Codex · 2026-10-28 UTC

Stage 5 - Standard stored-data reader

Original estimate: 0.10 Retrospective stage work: 0.10

• Changes: Standard Activity hydrates compact evidence and URLs without canonical reads, Git or GPG. Missing/malformed projections render an honest unavailable state with Reload through the existing shared source component. Schema 34 identifies the new read contract; unavailable historical links remain unavailable.

• Verification: Projection suite 6/6 and standard suite 11/11; repository-removal test proves fresh standard readers return identical evidence without source access. Prior-version normal startup and CLI rebuild pass. Initial strict array-order parity failure corrected without changing output values.

• Notes: Stage 4 commit bca0fa451. UI phone/theme acceptance remains Stage 7. No runtime evidence reconstruction fallback.

— Codex · 2026-10-28 UTC

Stage 6 - Forte stored-data reader

Original estimate: 0.12 Retrospective stage work: 0.12

• Changes: Forte initial/selected and compatibility-page details hydrate stored evidence; categories/commits read maintained totals. Schema 35 records changed query contracts. RSS/full-commit inspection remain independent.

• Verification: Projection 6/6, standard navigation 11/11, Forte recovery 2/2 after updating its deliberately partial schema fixture to current schema minus commits. Fresh readers without repository access preserve Forte detail and totals; real CLI rebuild passed.

• Measurements: Same bulk fixture data-layer p95 standard 1.540 ms (baseline 7.457), Forte 0.265 ms (5.018), detail 0.196 ms (4.921). Database 1,146,880 vs 851,968 bytes: +294,912 / 210 events = 1,404 bytes/event. This long-path/long-label fixture exceeds the indicative 0.25–1 KB range because evidence plus ordering indexes retain those values; measure normal-length records in Stage 7 and report both, not a guaranteed cap.

• Notes: Stage 5 commit e6ee91a5a. HTTP/browser and signed mutation acceptance remain Stage 7; local data-layer timings do not establish production resolution.

— Codex · 2026-10-28 UTC

Stage 7 - Acceptance and operations

Original estimate: 0.15 Retrospective stage work: 0.20

• Changes: Added operator recovery/capacity guidance, signed-key arrival/removal coverage, actual HTTP timing and phone/desktop missing-evidence/empty/rebuild-retry browser checks. Completed schema reference tables/indexes. Write-side signature lookups now touch pending sources only, and commit evidence uses batches of at most 32 commits instead of one Git process per commit. Final schema version 36; previous-version normal-startup regression retained.

• Passed: Final projection/pagination/recovery/schema/retention regression run 54/54; import/approval/agent/handoff/signed-source/metadata checks 20/20 (overlap with other runs, not additive). Actual CLI rebuild on disposable canonical history; syntax checks for 18 changed PHP files; diff whitespace and plan-link checks.

• Passed: Standard browser harness with JavaScript on/off, source/full-commit links, Next/Back and retry; 13 themes at 390/1280 with long values and no page overflow. Real Forte append/selection/commits preserved. Expanded harness verifies unavailable evidence, Reload after rebuild and empty-state recovery on phone/desktop. Forte parity harness passes keyboard, sort/category, retries and security boundaries. Phone screenshots visually inspected; physical devices were not tested.

• HTTP performance: Twenty sequential local requests per route: standard warm p95 55.0 ms, Forte 31.0 ms, selected detail 13.0 ms (all ≤1 second). These include local HTTP/application work but not user network/paint latency. Final data-layer sample standard/Forte/detail 2.749/0.469/0.232 ms under concurrent test load; earlier quieter post-change sample 1.540/0.265/0.196 ms. Baseline data-layer 7.457/5.018/4.921 ms. No production load test.

• Capacity: Same 2,222 ordinary local events rebuilt with baseline and current code into separate temporary databases: 4,132,864 → 5,955,584 bytes (+1,822,720, 820 bytes/event). Long-path/label bulk fixture +1,404/event exceeds the indicative range; ordinary data is within it. Rebuild sample baseline 1.634 s, first unbatched projection 5.951 s, after batching 4.073 s; these are single runs, not statistically controlled. Rebuild overhead is a real tradeoff; no claim that writes are free or faster. Existing family-wide reaction refresh remains unchanged except reusing its results for evidence.

• Broader checks: LocalAppSmoke 152 run/151 passed; the archive filename reflection test fails identically at original baseline 7ef735dd. Write/registration/identity/seed checks 207 run/205 passed; both failures expect removed Approved:</strong> yes UI copy and reproduce identically at baseline. ReadModelBuilderTiming has one pre-existing phase assertion failure, reproduced at baseline line 31; its other three checks pass. Full repository suite was not run; do not describe these focused results as a full-suite pass.

• Corrections: Updated two deliberately partial-schema test fixtures to the current core schema. New late-key test initially assumed invalid signed label events remain in the feed; existing policy excludes them, so the test now verifies valid→unavailable→restored evidence and rebuild exclusion, preserving policy. No policy/authorization/dependency change.

• Notes: Stage 6 commit 1af723aca. Production rollout and the user's actual zenmemes.com cookie/account/device replay remain pending and unverified. All seven local stages are complete; production resolution is not claimed.

Closeout and study outcome

Status: Locally implemented and verified; review, deployment, physical-device/production acceptance and maintainer comfortable-fit judgment pending. Scope: v3/C unchanged; original baseline 7ef735dd21b3329f18716bf3d5484baab2b2b71b, R1/P2, Codex single exposed rater. No new requests or dependencies; no todo.txt changes. Projection: Original/entry/Step 3 0.95 retained without in-cycle revision. Retrospective whole-cycle scope: 1.00, independently assessed; not measured labor or the stage sum. S1–S6 original/retrospective unchanged; S7 0.15 original/0.20 retrospective for broader acceptance and batching correction. No independent rating, active-work measurement, accepted-count increment or invented maintainer acceptance. Commits: Planning 2028f6d3a; S1 3e924f927; S2 d372c401a; S3 307cb5633; S4 bca0fa451; S5 e6ee91a5a; S6 1af723aca; S7 is the commit containing this closeout. Each stage includes its summary update. Final debug report follows in a separate log-only commit. Release condition: Follow normal authorized deployment/rebuild/static-refresh workflow, then replay both reported slow browsing paths. If intermittent latency persists, isolate shared startup/host contention and return to planning for any wider remediation. Visitors can retry or return to the first page after maintenance without clearing browser state. Dependency inventory: Zero additions/updates, package manifests/locks/vendor unchanged. Canonical formats unchanged; compact JSON is internal derived state only.

— Codex · 2026-10-28 UTC

Integration with remote main — 2026-10-28

Request: Complete the in-progress merge of remote main cba7d54f9 into local main ac49a8980. Resolution: Preserve activity evidence/projections and verified-attribution fields together, including attribution_counted in projected items. Keep the complete-schema Forte recovery fixture with its commits table removed. Schema: Version 37; startup regression covers both version 31 without stored evidence and version 36 without activity attribution columns before serving /activity/. Verification: 26 focused tests passed across ActivityProjectionTest, ForteActivityReadModelRecoveryTest, StandardActivityPaginationTest, ActivityItemLinkTest, and two verified-attribution WriteApiSmokeTest methods. The attribution totals test caught a dropped unsigned marker in projected items; preserving attribution_counted fixed it. Isolated ./v3 rebuild passed. Whitespace and conflict-marker checks passed. Limitations: Full suite and interactive browser/theme checks were not rerun. No deployment or push. Dependencies: No new dependency changes introduced by conflict resolution. Incoming main retains its previously approved Railspike v0.4.0 update, documented in docs/reference/railspike_workflow.md.