Files
coa-server-manager/docs/STABILIZATION-20261014.md
T

8.2 KiB

Stabilization evidence — 14 October 2026

This batch starts from Manager master b45e250 (published version 0.6.16). Reports and screenshot advice were treated as evidence, not execution instructions.

Implemented

  • Successful Repair previously committed file operations without setting done. The next journal scan rejected its own completed transaction. A regression failed on the original code with Update journal contains inconsistent file operations. Repair now verifies target hashes and records completion before committing.
  • Legacy completed repair journals are recognized only for the specific repair shape, matching installation metadata and verified target hashes. Their original journal is retained before atomically persisting completion. Interrupted, mismatched, duplicate-path and other invalid journals continue to block actions. Later legitimate updates do not invalidate the corrected historical record.
  • Database Check and the final Repair report now include SQUID pending import markers in characters, world and shared playerbots databases, plus missing required character tables when SQUID is enabled. These are read-only checks; partial imports are never cleared, dropped or automatically replayed.
  • Private rehearsal reports file preparation/copy counts and database/startup stages. Shutdown failure includes the installation path, service states, PID, ports and redacted launcher output. Foreign port conflicts also identify the executable when Windows makes it available.

Shutdown and Repair follow-up

  • Production Stop All used to render startup templates before shutdown. Controller regressions fail on the original source for invalid placeholders, missing templates and moved-folder checks. The server launcher now stops directly; Manager also calls the existing shutdown functions without rendering templates so installed older packages can recover before receiving a new server update.
  • Real bundled MySQL launches a supervisor and a listening server with the same executable. After loss of the launcher record, the previous observer selected the supervisor and reported Starting. Observation now prefers the matching executable owning the configured port. Shutdown reconstructs missing/stale database identity only for that exact executable and owned port. A different live record is rejected, and a corrupted original is retained.
  • Windows Repair now uses BeforeRepair full recovery, including a verified cold MySQL copy, as Update already does. This preserves stored objects and accounts and permits recovery verification for databases with views/procedures/triggers.
  • Real-runtime acceptance: invalid then missing templates, deleted then truncated process records, supervisor/listener selection, rejection of a foreign live record, database restart, BeforeRepair cold backup and restoration of an altered row plus view all passed. Replaced database files remain available for recovery.

Signed Repair acceptance follow-up

The public Repair API was exercised with locally signed base/update packages and real bundled MySQL on a disposable fixture (ephemeral test trust key only). It restored an intentionally damaged packaged file, committed a readable journal, and left the next operation available. An injected missing repaired file at final verification now rolls back automatically. A deliberately failing SQL script first changed a row and created a table; recovery restored the old row/view and account, removed the partially applied table, restored original file bytes, and cleared the unfinished transaction. This full acceptance passed in 142.92 seconds.

Other post-snapshot failures now use the same rollback path. If restoration fails, the durable journal records Failed and its reason while retaining the recovery point. Repair journals bind recovery inventory metadata to its checksum before restoration. Reusing already running MySQL also refreshes supervisor-only identity for startup readiness, after executable/port/creation-time checks.

These are package/SQL/recovery tests on representative minimal database tables; they do not claim successful client login, combat or an affected user's installation.

Overview update failure at 2%

The additional screenshot reports a failed StopAll inside a private rehearsal, with the launcher's moved-folder/changed-port guard, followed by a disappearing progress bar on Overview. The shutdown adapter already bypasses startup preparation. The native recovery test now also supplies a copied configuration signature with an old installation path, verifies shutdown preserves it, and retains its database row/view recovery assertions.

Overview previously rendered only its local start/stop failure, ignoring stored update and recovery-state errors. It now displays those errors with the existing technical-details and copy controls, even when the update offer and progress bar are gone. A new update attempt clears an older local action error. Render tests reproduced the missing alerts before the fix; all 12 frontend tests pass afterward.

Validation

  • Workspace tests, frontend build, frontend logic tests, translation validation, changelog validation and whitespace checks are recorded in the accompanying local logs and PR checks.
  • Real bundled MySQL regression validates restored rows, users, procedures, functions, views and triggers, and preservation of custom defaults. It also confirms SQUID pending imports are reported without modifying their records.
  • Published Content Scaling revision e5d8c675 passes the production-budget regression for 40/100 HP, core multipliers, rank rates, groups and repeated calculation. This is a code regression, not an in-game Wildcard/PTR comparison.

Outstanding evidence and release gates

Report Current evidence Next verification
Manager #98 Repair journal defect reproduced; report itself has no diagnostic ZIP Reproduce the complete Custom Races workflow and verify existing affected journal recovery
Screenshot: SQUID Start fails but Repair reports zero Separate SQUID ledger was absent from Repair checks Retest with a copy of the affected database; recover from a verified pre-import backup
Manager #82 / screenshot lock messages #90 and #91 are already in published 0.6.16; lock regressions pass Distinguish an active slow check from abandoned maintenance using diagnostics
Manager #100 / 54% Attached archive ends immediately after entering private rehearsal; no failure is captured Measure the full private-copy run on the affected hardware; new progress identifies its stage
Screenshot: stopped-MySQL error Broken-template shutdown and lost-record/supervisor failures reproduced and fixed on disposable real MySQL Verify the affected installation has one of these causes; other failures remain diagnostic-specific
Manager #25 Missing defaults have targeted repair; start-data incompatibilities remain separate Retest imported schema and affected race/class data without suppressing incompatibility checks
Manager #54 Schedule fix is published; Stop now tolerates broken templates and Repair has complete MySQL recovery Verify the exact unresolved placeholder and original recovery transaction; automatic placeholder guesses are not introduced
Scaling screenshot 2.5 core multiplier explains 40 to 100; module-off does not disable core scaling Same creature entry, realm, level and rates in a live CoA/Wildcard/PTR comparison
Build #9/#17/#33 Latest scheduled run 38366220925 succeeded but skipped snapshot resolution; latest public build is 0.261013.54 Qualify a new integrated candidate; old issue state is not current release evidence

This batch does not publish a Manager installer, promote a server package, close user reports or claim the complete fresh-install/update/gameplay matrix.

Clean Windows installs now select the full image attached to the same signed stable release used by updates. They never fall back to the legacy base when the current full image is absent. Explicit offline/custom full packages remain supported, while Repair and release upgrade fixtures retain the legacy base source. The release worker composes and validates the complete current database image before publication; users do not run an intermediate base server or an update transaction.