Reduce repeated full-projection lookups during workflow execution

[rmcdaniel]

Problem

Relates to #508. Repeated full monitoring projection rereads each existing timeline, wait, and timer row individually. Summary projection also derives activity and timer views more than once in the same pass. A small embedded workflow with three parallel activities followed by a timer per round exposes the growing cost.

Changes

• Prefetch each run's existing projection rows using the configured model, table, and connection; reuse those Eloquent instances within that pass only.
• Retain casts, model save hooks, duplicate-key recovery, point-delete stale cleanup, and history-based repair. Incremental child/claim projection paths are unchanged.
• Reuse the activity metadata and timer views already computed by full summary projection.
• Add fixed-query-count, unchanged-row/timestamp, repair, unrelated-run, custom-model/connection, and cold-reload workflow regression tests.

No new cache, background projector, schema, durable-history format, or public execution contract.

Validation

Local PHP 8.4.25 / Laravel 13 source checks:

• 15 focused tests / 98 assertions passed, including cold activity metadata projection without external-payload reads and configured-model/table/secondary-connection cases.
• Full phpstan analyse src tests and changed-file ECS passed.
• Full source CI passed at c59828f4c7be77c6efd078419d0aed548c26c8fe: unit/coverage, all MySQL and PostgreSQL shards, MariaDB, Laravel 9-13 upgrade checks, regression corpus, contracts, and quality. PR checks also passed.
• Final source review confirmed that prefetch is run-scoped and same-pass only; incremental child/claim paths, configured model connections, duplicate-key handling, and stale point-delete behavior are preserved.

Diagnostic only, not a published capacity benchmark: matched sequential runs on the same host, container limit 1 CPU / 1 GiB, SQLite in memory, fake no-op activity bodies, Workflow 2.0.12 versus this source overlay. Sixteen rounds produced 48 activity completions and 179 history events; both runs completed with the correct output and activity count.

Measurement	Baseline	Candidate
Total test duration	83.391 s	65.224 s
Round 15 duration	10.050 s	7.929 s
Round 15 SQL queries	5,860	1,196
PHP peak allocated memory	79 MB	79 MB

XHProf on a separate three-round baseline attributes about 75% of elapsed profiled time to full summary projection. Profiler overhead is excluded from the timing table. The repository-owned three-round correctness reproduction is tests/Unit/V2/ProjectionReplayTest.php; query-count regression is in ProjectionPrefetchTest.php.

This is a partial optimization: the remaining history-dependent PHP cost is still material. Keep #508 open until its acceptance criteria, broader correctness checks, and published-consumer follow-through are complete. No claim that this patch alone resolves the full performance problem.

Before Delivery

• Source checks and review are complete. Publish the patch and verify installed-package behavior before calling it delivered.
• Keep remaining history-dependent work explicit; this patch removes measured redundant work without claiming constant-time full history projection.
• On publication, audit pinned Workflow consumers (including Server) and run affected conformance with published artifacts. Do not release unchanged language SDKs merely to align version numbers.

Current main consumer audit: Server requires Workflow 2.0.12 exactly and will need a dependency update and rebuilt release. Sample App's embedded lock is 2.0.12 under ^2.0, requiring a lock update and first-user verification. Waterline's Workflow dependency is development-only; refresh relevant embedded integration qualification without claiming a runtime package release is necessary. PHP SDK and CLI have no Workflow runtime dependency; there is no portable wire change requiring synchronized SDK releases. Deployment-specific follow-through stays in its owning private record.

[rmcdaniel]

The source change passed full local PHPStan and 15 focused cases in aggregate, covering configured model/table/connection routing, cold-reload projection consistency, metadata-only external payload behavior, existing unique-key recovery, and fixed projection read counts. PR functional/corpus checks passed; the first style check caught one chained-call indentation introduced by the formatter. Commit c59828f corrects it and a second local style check is clean. The first broad run was cancelled and one complete source validation has been requested for the corrected commit. This remains draft, with no published-consumer or production performance claim.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.06%. Comparing base (3994044) to head (c59828f).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #509      +/-   ##
==========================================
+ Coverage   88.04%   88.06%   +0.02%     
==========================================
  Files         460      460              
  Lines       63036    63059      +23     
==========================================
+ Hits        55497    55531      +34     
+ Misses       7539     7528      -11     

Flag	Coverage Δ
v2	88.06% <100.00%> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.