perf(core): spread spray folder scans over timer ticks

[Rushaway]

Fixes #36

@cmer81 could you review this? It removes the 89 ms OnMapEnd and 57 ms TimerNext hitches by spreading the spray folder work over timer ticks.

What changed

• OnMapEnd does no file work. It only starts a background scan.
• One scan job, limited per tick. SprayScan replaces RecursiveFiles, RecursiveSearchDirs, TimerNext and DeleteEmptyDirs. It runs on a 0.1 s repeating timer (TimerScan) and stops after SCAN_BUDGET (2 ms) each tick, then continues where it left off. It runs in three steps:
  1. Walk the folders one entry at a time, keeping the directory listing open between ticks.
  2. Check (sm_spray_test) or move (map end, plugin start) the queued sprays.
  3. Remove emptied folders, deepest first. RemoveDir fails on folders that still hold files, so only empty ones go.
• Plain filesystem calls. Scan file calls no longer use Valve FS, which searches every game path on each call. FileCheck already reads sprays this way.
• No list shifting. Queued files are read by index instead of Erase(0).
• Backup folder fix. The scan creates <root>/backup_sprays, the folder sprays are actually moved into. On CS:S, the old code created download/user_custom/backup_sprays instead.
• sm_spray_test prints to the server console if the admin leaves before it finishes, instead of throwing an error.

OnPlayerRunCmd is unchanged. Outside impulse 202 it does one integer comparison, and removing it would reopen the jingle exploit.

Behaviour changes

• If a move is still running when the next map ends, a new one isn't started. Files it misses are moved at the following map end.
• A very large folder takes longer in wall-clock time. The old code also kept moving sprays into the next map, so this isn't new, but it can last longer. Raise SCAN_BUDGET to speed it up.

Testing

• Builds with SourceMod 1.12.0.7255, no warnings
• Change map: TimerScan stays around 2 ms and OnMapEnd drops out of the profiler
• download/backup_sprays fills up and emptied user_custom/xx folders are removed
• sm_spray_test finishes and prints its summary

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🟡 Changes recommended

Backup-directory creation still uses Valve FS, contrary to the performance objective.

Review effort: Balanced
Findings: 1

Open (1)

• Avoid Valve FS when creating the backup directory · New

What changed in this PR

Refactors spray-folder maintenance into incremental timer-based scans to reduce server-thread stalls.

Changes:

• Adds budgeted asynchronous scanning, validation, moving, and cleanup.
• Fixes backup-folder targeting and disconnected-admin output.
• Updates Copilot guidance for the new scanning architecture.

File	Description
FixSprayExploit.sp	Implements incremental spray scanning and cleanup.
.github/​copilot-instructions.md	Documents the updated workflow.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[cmer81]

Reviewed and built on SourceMod 1.13.0.7427 (base and head compile, no new warnings).

The incremental scan looks correct: directories are removed leaf-first (visited is filled parent-before-child and popped in reverse, RemoveDir refuses non-empty folders), the scan root is no longer a removal candidate, the active guard prevents overlapping scans and all handles are freed at the end, and a disconnected admin falls back to the server console.

Performance: OnMapEnd now only queues the scan, and each tick does at most ~2 ms of work with the overshoot bounded to a single file operation (the clock is checked before every operation). Replacing Erase(0) by an index also removes the O(n^2) cost on large queues. The trade-off is a throughput cap (~2% duty cycle), which is acceptable here. This is from reading the code, I have not profiled it on a live server.

One thing to fix before merging: the behaviour changes but PLUGIN_VERSION is still "2.28", please bump it (e.g. 2.29).

Minor, non-blocking: the final sm_spray_test message always goes to the console now (the original reply source is not kept), which is fine.