PHP: lsp_cross/parallel.resolve livelocks + unbounded RAM on large codebases (no PHP step-budget guard) — public repro: Shopware

[ilya-kudravets]

Version: codebase-memory-mcp v0.9.0 (macOS arm64, release binary)
Language: PHP
Mode: reproduces in fast and moderate.

Summary

On large PHP codebases, extraction and the definitions pass complete normally, then pass=lsp_cross → parallel.resolve.start ... workers=15 never finishes. With the supervisor on it is reaped as outcome=hang signal=9; with it off it runs indefinitely.

It is not a clean deadlock and not steady progress — a livelock with unbounded memory churn: an initial ~99% CPU burst spikes RSS to multiple GB, then CPU collapses to ~10–18% while RSS thrashes up and down for the rest of the run, making no forward progress. Only CBM_DISABLE_LSP_CROSS=1 avoids it.

Public reproduction (Shopware)

git clone --depth 1 https://github.com/shopware/platform.git   # tested @ 65110f78, ~19.5k php files
codebase-memory-mcp cli index_repository '{"repo_path":"./platform","mode":"fast","name":"sw"}'

Result — hangs. Extraction fine (parallel.extract.done nodes=171136), then parallel.resolve.start files=19483 workers=15 and no parallel.resolve.done. Worker sampled during resolve:

t1: CPU=18%  RSS=2811MB      t5: CPU=14%  RSS=2802MB
t2: CPU=13%  RSS=3246MB      t6: CPU=97%  RSS=2068MB
t3: CPU=13%  RSS=1923MB      t7: CPU=96%  RSS=1116MB
t4: CPU=13%  RSS=2847MB      t8: CPU=17%  RSS=2804MB   (never completes; reaped outcome=hang)

Counter-check — same repo with CBM_DISABLE_LSP_CROSS=1 completes cleanly: status=indexed, 160,945 nodes / 530,740 edges. So the hang is isolated to the cross-file resolution pass.

(Independently reproduced on a private Symfony 5.4 / API Platform 2.6 app, ~3,700 files, identical signature — and on a framework-free php-only src/ copy with no vendor/config present, so it is not a vendor/framework artifact.)

It scales with cross-module reference density, not one bad file

Delta-debugging the private app (clean folders, php-only, fast): no single top-level directory hangs, and neither half of the source tree hangs — only the whole tree does. The pathology forms across the module graph and needs the full cross-reference density, consistent with a cyclic/superlinear type-resolution walk rather than a single pathological construct. (This is also why splitting a repo to minimize it makes the hang vanish.)

Root-cause signal: the PHP resolver has no step/parse budget

The C, TS, and Python resolvers self-limit and log it — [clsp] expression eval step budget exhausted, [tslsp] type-text parse budget exhausted, env CBM_TS_TYPE_BUDGET — and this exact pass's deadlocks were already fixed for those languages:

• cbm_pipeline_pass_lsp_cross deadlocks on a specific .ts file (no progress, ~3 CPU-s, indefinite) #370 (closed) — cbm_pipeline_pass_lsp_cross deadlocks on a specific .ts file (no progress, ~3 CPU-s, indefinite) — fixed for TypeScript.
• py_lsp: py_eval_expr_type infinite recursion — Python missed in 9f981a1 depth-guard fix #720 (closed) — py_eval_expr_type infinite recursion, fixed with a depth guard; Python parser hangs on deeply chained method calls during definitions pass #710 (closed) — Python parser hangs on deeply chained method calls — both Python hang fixes in the same resolution machinery.

There is no equivalent [php_lsp] budget, and CBM_LSP_MAX_WALK_DEPTH=3 does not help — so the PHP path appears never to have received the step/recursion budget the other resolvers got.

Two issues

1. PHP lsp_cross livelock / unbounded RAM — port the step/recursion/cycle budget from the TS (cbm_pipeline_pass_lsp_cross deadlocks on a specific .ts file (no progress, ~3 CPU-s, indefinite) #370) and Python (py_lsp: py_eval_expr_type infinite recursion — Python missed in 9f981a1 depth-guard fix #720/Python parser hangs on deeply chained method calls during definitions pass #710) fixes to the PHP type resolver, so cross-module resolution degrades (drop the edge / mark unresolved) instead of churning forever.
2. Supervisor gap — a hang in the lsp_cross pass (not per-file extraction) has no file to quarantine, so CBM_INDEX_SUPERVISOR just SIGKILLs and restart-loops through CBM_INDEX_MAX_RESTARTS with zero forward progress. A pass-level timeout that skips/degrades cross-resolution would fail gracefully. (Relatedly, CBM_MEM_BUDGET_MB is not enforced as a hard cap here — RSS ran well past it.)