read_model_hot_indexes — Railspike debug log

2026-10-24 UTC — Step 1 (Solution Assessment)

Necessary/useful instructions

• RAILSPIKE.md "Step Guide" + step1_solution_assessment.md structure contract (Original Query → Understood Intent → Problem → ≥2 options → Recommendation): kept the assessment to one page and forced an explicit isolation rationale for the schema-change cycle. Retain.

• CONTRIBUTING.md read-model schema-change checklist (referenced, not part of Railspike): materially shaped the options and the verification list. Retain; this cross-reference is the reason Option A is scoped as it is.

• Effort study fields (AGENTS.md standing P2 enrollment + ledger R1 recovery note): enabled a concrete projected total (≈0.30) instead of a deferred "unavailable" because R1 frozen references were recovered 2026-10-23. Useful; retain.

Redundant/unnecessary instructions

• No supported findings. (The effort-estimation guidance is long, but each clause consulted was load-bearing for this cycle's snapshot.)

Inapplicable instructions

• Extended estimates (docs/effort-estimation.md#extended-estimates): not enabled for this cycle; no per-option estimate line required. Non-use is not redundancy.

• Follow-up-cycle / branch-reuse rules: no Step 4 branch exists yet.

Uncertain assessments

• Projected total ≈0.30 is a single exposed, non-blind rater judgment against R1; confidence is moderate because the schema-change checklist cost is well-characterized but rebuild time on a large instance is unmeasured here.

2026-10-24 UTC — Step 2 (Feature Description)

Necessary/useful instructions

• step2_feature_description.md "Known user or environment" clause: correctly prompted an explicit omission (no named user/env) rather than silently dropping the section. Retain — the explicit-omission guidance prevented an ambiguous gap.

• "A feature is a normal-flow vertical slice, not a component, API, cache, migration" guardrail: forced framing an inherently infrastructure change (indexes) as reader/operator outcomes with entry→outcome→recovery, which kept the completion boundary user-facing. Retain.

• Shared component inventory requirement: surfaced that PublicOfflineSnapshotBuilder already defines the equivalent index, reinforcing the "bring main model in line" framing and confirming no new component. Retain.

Redundant/unnecessary instructions

• No supported findings.

Inapplicable instructions

• Documentation-only scope rules: this is an application change; recorded the work type explicitly instead.

• Extended estimates: not enabled; no per-item estimate annotation required.

Uncertain assessments

• Rebuild cost on a large production instance is unmeasured in this environment; flagged as a risk with ./v3 rebuild as earliest validation rather than asserted as negligible.

• ≈0.30 reaffirmed is a single exposed, non-blind rater judgment; confidence moderate because the schema-change checklist is well-characterized.

2026-10-24 UTC — Step 3 (Development Plan)

Necessary/useful instructions

• step3_development_plan_do.md per-stage fields (Goal/Dependencies/Expected changes/Verification/Risks/Canonical components): surfaced a concrete, otherwise-easy-to-miss dependency — ReadModelSchemaTest asserts the exact table/index name set, so it must change in the same stage as the index additions. Retain; the structured fields caught a real coupling.

• ## Key Risks near the top with **High risk:** prefix: forced the results-drift risk to the top where a reviewer sees it first. Retain.

• "Document database changes conceptually (no SQL)" guardrail: kept the plan to index names + column lists rather than DDL, which stayed readable while still unambiguous. Retain.

• Completion Contract "required recovery" field: prompted tracing the actual rebuild trigger (Application::ensureReadModel() → schema_identity_mismatch), making Stage 3's regression test concrete rather than hand-wavy. Retain.

Redundant/unnecessary instructions

• No supported findings.

Inapplicable instructions

• Extended-estimate stage-work bullets: mode not enabled; per-stage estimate bullet omitted by rule.

• Known-user/environment replay stage: Step 2 named no user/environment; no replay stage required.

• Documentation-only file-scope checks: this is an application change.

Uncertain assessments

• Stage 4 is verification-only; whether it should be folded into Stage 3 is a judgment call — kept separate so query-plan/content evidence has a distinct, auditable handoff. Low-confidence either way.

• ≈0.30 reaffirmed is a single exposed, non-blind rater judgment; rebuild timing on a large instance remains the main unmodeled factor.

2026-10-24 UTC — Step 3 (Development Plan) — Revision: extended estimates enabled

Revision of the 2026-10-24 Step 3 entry above (does not supersede its findings; adds the extended-mode observation). User requested enabling extended estimates for this cycle; Step 3 re-drafted with per-stage stage-work values. Step 3 approval remains pending; this revision does not advance the step.

Necessary/useful instructions

• RAILSPIKE.md extended-estimate rules ("every Step 3 stage gets a separately labeled stage work estimate"; "never sum stage estimates"; "cycle total independently judged"): kept the four stage annotations (0.05/0.07/0.15/0.08) distinct from the reaffirmed ≈0.30 cycle total and prevented presenting their sum as the projection. Retain — the explicit no-sum rule is load-bearing here.

• "do not backfill missing historical estimates" + "record an explicit disable/enable in the current artifact": correctly drove recording the mode as effective from Step 3 and NOT inventing retroactive Step 1 per-option estimates. Retain.

• "identify the owner of shared work without charging whole-cycle planning to every stage": made Stage 4 the explicit owner of end-to-end/deployment verification, avoiding double-counting that overhead across stages. Retain.

Redundant/unnecessary instructions

• No supported findings.

Inapplicable instructions

• Documentation-only file-scope checks: application change.

• Known-user/environment replay stage: none named.

Uncertain assessments

• Stage 3's ≈0.15 is the least certain stage value: it depends on how cleanly Application::ensureReadModel() can be driven from a test harness without heavy fixtures; flagged as medium uncertainty in-plan.

• The per-stage values are single exposed, non-blind rater judgments against R1; they are annotations, not measured effort.

2026-10-24 UTC — Step 4 (Implementation) — completed-step report

Single Step 4 report (not per stage). Stages 1–4 implemented and committed on feat/read-model-hot-indexes; planning 7f928218, stages 4eb59792/87fceacd/cef85b1d/3a25919c.

Necessary/useful instructions

• Stage Boundary Protocol (finish → verify → update summary → git status → stage commit → git log → next): kept each stage independently committed with its own verification, and caught that the two unrelated debug logs must stay unstaged at every boundary. Retain — the explicit git status gate prevented cross-cycle leakage.

• "A focused pass is never reported as a full-suite result" + "identify reproduced baseline failures": directly shaped the honest Stage 4 reporting of the two pre-existing ReadModelProvenanceTest git failures as reproduced-baseline, not regressions. Retain — load-bearing for an accurate handoff.

• Extended-mode per-stage retrospective requirement: surfaced useful signal (Stage 3 came in under estimate because existing fixtures/public API avoided a heavier harness; Stage 4 ran over due to environmental PHP linkage). Retain.

• CONTRIBUTING.md schema-change checklist (external to Railspike but referenced throughout): every clause mapped to a concrete stage/verification. Retain the cross-reference.

Redundant/unnecessary instructions

• No supported findings.

Inapplicable instructions

• Documentation-only verification gates (git diff --check, allowed-file-scope): this was an application change; used runtime/test verification instead.

• Known-user/environment recovery replay: none named in Step 2.

• Browser/UI state matrix: no presentation change (indexes only; content proven byte-identical), recorded not-applicable with reason.

Uncertain assessments

• The ≈0.30 cycle total and per-stage retrospectives are single exposed, non-blind rater judgments; not measured effort. Maintainer comfortable-fit judgment is pending (unreviewed/unmerged), so study outcome acceptance stays pending — explicitly not invented.

• Whether Stage 4 should have been folded into Stage 3 remains a low-confidence call; keeping it separate gave the query-plan/rebuild evidence a clean auditable commit.