[taylorg009]

Found by the ContextBench run (#201). Indexing element-hq/element-web@c9d9c421 never finished: the build ran 2 h 9 min before it was killed. A per-file scan pinned it on docs/lib/mermaid.min.js (2,935,907 bytes, lines up to 278,302 characters). A daemon opened on any repository that vendors a bundle like this would hang the same way.

Cause (gdb samples, release build)

The time was in the local-value analysis behind HTTP-client detection (http_call_handler -> analyze_client_call -> collect_url_template / options_method -> local_values).

• Scope walk per read. local_values walked the variable's whole scope for every read, via visit_named_descendants.
• Block rescan per lookup. local_binding rescanned the enclosing block's statements on every lookup.
• Quadratic child access. Both walked children with ts_node_named_child(node, i), which costs O(i).

In a minified bundle the scope is most of the file and reads number in the thousands.

Fix (src/engine/http_consumers.cpp)

• ExtractionMemo. A per-parse memo (language_config.hpp; extractor.cpp gives each parse a fresh one, because it holds that parse's tree nodes) indexes, once:

  • each scope's writes, by the name each write can affect, in document order;
  • each block's first declarator of each name.

  local_values and local_binding read these indexes. The per-write logic is unchanged.

• Cursor walks. visit_named_descendants and the block scan use a tree cursor instead of indexed child lookups.

• One pattern traversal. binds_pattern is now a lookup over pattern_names, the same traversal the write index uses.

Evidence

	Before	After
mermaid.min.js alone	killed after 60 s, 5 min and 2 h	17-28 s
Synthetic, N = 500	1.74 s	0.26 s
Synthetic, N = 1500	14.2 s	0.23 s
New scale test (N = 3000)	62.9 s, fails	0.15 s, passes

Synthetic N means N let locals, N unrelated writes and N fetch(local) calls in one function.

• Same output: the original code never finished mermaid.min.js, so its graph is compared across the steps of this fix instead: the first step (memo only, 51.5 s) and the final code produce identical graphs (8,318 nodes, 180,429 links, node ids and links compared in full). Against the original code, equality is shown on the synthetic inputs. The synthetic outputs are link-for-link equal, with all N calls resolved to /api/v1/items/<i>.
• Tests: full ctest (release, mars) 99/100. The one failure is cgraph_file_watcher_test, the known mars-only failure.
• Not addressed here: the remaining 17-28 s on that file is spread across ordinary pipeline stages (community layout, exports, call resolution). Whether minified bundles should be indexed at all is a separate policy question.

🤖 Generated with Claude Code

[github-actions]

PR brief: 0 agent commits · run the full suite

Warning

Run the full suite. The graph cannot vouch for this diff, so every test file is selected. Nothing is skipped; this is the safe default, not a failure.

8 files have no graph node (config, asset, or unextracted).

If these files cannot change which tests should run, add a pattern for them to ignore. If they feed tests (fixtures, data), this outcome is correct.

bench/billed-cost/results  ████████████████████████  7
bench/billed-cost          ███  1

All 8 unmapped files

bench/billed-cost/results/multi-cgraph.json
bench/billed-cost/results/multi-cgraph_directed.json
bench/billed-cost/results/multi-gold-recall.json
bench/billed-cost/results/multi-hook-live-cgraph.json
bench/billed-cost/results/multi-hook-live-gold-recall.json
bench/billed-cost/results/multi-read-hook-cgraph.json
bench/billed-cost/results/multi-read-hook-gold-recall.json
bench/billed-cost/tasks-multi.json

Computed at a225121 against base 43453f3.

Summary	PR #204
Intent	no checkpoints on this branch (6 commits)
Since push 7e1057d	1 new commit, -5 changed files, -49 tests reached

What each commit did

Commit	Intent	Files	Reaches	Ran before push
52b7746 fix(js): a minified bundle no longer takes hours to extract	no checkpoint	4 files
7e1057d test(client): wait for the lone repository's graph before a…	no checkpoint	1 file
1403ab0 ci: cold-install timing on clean macOS/Linux runners + Home…	no checkpoint	5 files
51105f8 test(client): wait for the lone repository's graph before a…	no checkpoint
c106d73 bench: multi-file billed-cost A/B — agents still don't call…	no checkpoint	11 files
a225121 style(js): put visit_named_descendants' comment back above …	no checkpoint	1 file

Claims checked

no checkpoint on this branch makes a claim the graph can check

_{blastline 0.13.0. Selection is advisory unless your workflow gates on it. How selection works. Intent: 0 of 6 commits. Not checked: symbol changes: no --change-context given.}