feat(confgen): parallelize per-sheet row parsing with bounded worker pool

[Kybxd]

Summary

Add an optional intra-sheet parallel parsing path for vertical map / list /
struct sheets. Total concurrently running row-parsing workers across all
sheets are capped at GOMAXPROCS via a process-wide counting semaphore, so
the existing book/sheet level fan-out can stay unbounded without inducing
scheduler thrash.

Motivation

Profiles on large workbooks (e.g. zonesvr's 580k-row ActorConf) showed
single-sheet row parsing dominating the critical path. Naively spawning
workers per sheet would push live goroutine count to
N_concurrent_sheets * GOMAXPROCS, regressing per-sheet timings vs an
isolated single-sheet run. We need worker-level parallelism that does NOT
oversubscribe the scheduler.

Design

• Eligibility (parallel.go: parallelEligibility): walks the
  descriptor to decide whether a sheet's layout is safe to parallelize
  (vertical map/list/struct, no transpose, etc.). Cached only per call,
  not globally, because it depends on per-sheet WorksheetOptions.
• Planner (planShards): two strategies.
  • strategyRowRange: even row blocks, used when row order is the only
    invariant.
  • strategyKeyHash: pre-scan the outer-key column via the new
    tableparser.ScanColumn, FNV-1a-hash each cell to a bucket so all
    rows sharing an outer key go to the same worker (preserves
    last-write-wins / append-order / E2005 duplicate-key detection).
• Workers: shallow-cloned sheetParser with a private
  sheet-collector child and fresh cards map, parsing into per-shard
  partial messages that the main goroutine merges after g.Wait.
• Concurrency cap (xpool.NewCPUSemaphore): leaf workers acquire a
  token before the heavy parse phase. The semaphore is intentionally
  not shared with gen.convert (book/sheet level): an outer
  goroutine holding a token would deadlock waiting for its own inner
  workers.

Files

• internal/confgen/parallel.go (+ parallel_test.go): planner,
  worker fan-out, merge, eligibility walk, bucketForKey. The latter
  carries a documented risk note: it hashes raw cell strings, so
  variants of the same logical key ("1" vs "01", "true" vs
  "TRUE", etc.) can route to different buckets. Two follow-ups are
  spelled out in the comment for when this actually surfaces:
  1. hash the typed key by decoding via the outer-key field descriptor;
  2. add a merge-time uniqueness/order check by typed key as a
     defensive backstop.
• internal/x/xpool/sem.go (+ sem_test.go): a tiny context-aware
  counting semaphore (NewSemaphore, NewCPUSemaphore).
• internal/confgen/parser.go: parse/export sum-time logging in
  ScatterAndExport for perf diagnostics (CPU-style sums, not
  wall-clock).
• internal/confgen/table_parser.go,
  internal/importer/book/tableparser/parser.go: expose ScanColumn
  and rangeDataRowsAtIndices / rangeDataRowsInRange so the planner
  can pre-scan the outer-key column and workers can iterate
  non-contiguous row sets.
• test/functest/...ParallelConf...: functional fixture (sheet, proto,
  golden JSON).

Performance

zonesvr (135 workbooks, 351 sheets), N=5 full runs, identical binary,
only workerSem capacity differs:

mode	wall median	min	max	stdev
cap = GOMAXPROCS (kept)	13501ms	13237	13684	165
cap = 1<<20 (effectively uncapped)	13739ms	13657	13847	79

~1.76% wall improvement; more importantly, per-sheet timings of
critical-path large sheets under concurrent load are now on par with
their isolated single-sheet runs, instead of being amplified by
scheduler contention.

Tests

• New unit tests for xpool semaphore and bucketForKey /
  planShards.
• New functest fixture ParallelConf exercising the parallel path
  end-to-end.
• go test ./... green locally.

Known follow-ups

• bucketForKey string-vs-typed-key equivalence (documented inline,
  two fix paths listed; deferred until observed in real configs).
• Parse/export timing logs in ScatterAndExport are perf-diagnostic;
  consider gating behind a debug flag in a follow-up if too noisy.

[github-actions]

The latest Buf updates on your PR. Results from workflow Buf CI / buf (pull_request).

Build	Format	Lint	Breaking	Updated (UTC)
✅ passed	✅ passed	✅ passed	✅ passed	May 29, 2026, 1:27 PM

[codecov]

Codecov Report

❌ Patch coverage is 80.84416% with 59 lines in your changes missing coverage. Please review.
✅ Project coverage is 75.07%. Comparing base (6863556) to head (8f5fc13).

Files with missing lines	Patch %	Lines
internal/confgen/parallel.go	83.41%	18 Missing and 15 partials ⚠️
internal/importer/book/tableparser/parser.go	59.25%	14 Missing and 8 partials ⚠️
internal/confgen/parser.go	83.33%	2 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #411      +/-   ##
==========================================
+ Coverage   74.87%   75.07%   +0.20%     
==========================================
  Files          88       90       +2     
  Lines        9384     9688     +304     
==========================================
+ Hits         7026     7273     +247     
- Misses       1785     1818      +33     
- Partials      573      597      +24     

☔ View full report in Codecov by Sentry.
📢 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.