fix: address code review issues — parser-stage correctness, memory cap, O(N²) chunked fetch, and docs

[szibis]

Summary

• Fix 1 (HIGH correctness): Restore __error__ guard for parser-stage metric queries. VL stats_query_range counts parse-failed lines while Loki excludes them. All rate/bytes_rate/count_over_time/bytes_over_time queries with parser stages now route to the slow log-fetch path unless the caller explicitly adds | drop __error__ to opt in to VL's count-all semantics. The opt-in check uses the inner LogQL base query (without outer aggregation or range window brackets) so hasDropErrorOnlyPostParserStage correctly identifies the clause.
• Fix 2 (MEDIUM performance): Eliminate O(N²) copy amplification in coldBackwardChunkedFetch. Replaced the append(chunkBody, accumulated...) prepend loop with a slice of chunks reversed and joined once at the end.
• Fix 3 (MEDIUM memory): Cap unbounded io.ReadAll in collectRangeMetricHits at 64 MB via the existing readBodyLimited helper to bound RSS on oversized VL error response bodies.
• Fix 4 & 5 (documentation): Document optional tenant header behaviour and multi-tenant serial fanout in KNOWN_ISSUES.md. Add a TODO comment in multitenant.go referencing the tracking entry.

Test Plan

• go test ./internal/proxy/... passes (1638 tests, 0 failures)
• TestQueryRange_RateParserStageTumblingUsesSlowPath verifies parser-stage rate without | drop __error__ uses slow path
• TestQueryRange_RateParserStageDropErrTumblingUsesStatsQueryRange verifies | drop __error__ opt-in restores fast path
• TestQueryRange_CountOverTimeParserUsesDirectStatsRange verifies grouped count_over_time with | drop __error__, __error_details__ uses stats_query_range
• TestCompatHelpers_ParseQuantileAndUnwrapErrorName verifies updated unit assertions for parser-stage slow path

[github-actions]

PR Quality Report

Compared against base branch main.

Coverage and tests

Signal	Base	PR	Delta
Test count	2506	2507	1
Coverage	87.2%	87.2%	0.0% (stable)

Compatibility

Track	Base	PR	Delta
Loki API	100.0%	11/11 (100.0%)	0.0% (stable)
Logs Drilldown	100.0%	17/17 (100.0%)	0.0% (stable)
VictoriaLogs	100.0%	11/11 (100.0%)	0.0% (stable)

Performance smoke

Lower CPU cost (ns/op) is better. Lower benchmark memory cost (B/op, allocs/op) is better. Higher throughput is better. Lower load-test memory growth is better. Benchmark rows are medians from repeated samples.

Signal	Base	PR	Delta
QueryRange cache-hit CPU cost	1603.0 ns/op	1653.0 ns/op	+3.1% (stable)
QueryRange cache-hit memory	200.0 B/op	200.0 B/op	0.0% (stable)
QueryRange cache-hit allocations	7.0 allocs/op	7.0 allocs/op	0.0% (stable)
QueryRange cache-bypass CPU cost	1916.0 ns/op	1942.0 ns/op	+1.4% (stable)
QueryRange cache-bypass memory	281.0 B/op	282.0 B/op	+0.4% (stable)
QueryRange cache-bypass allocations	7.0 allocs/op	7.0 allocs/op	0.0% (stable)
Labels cache-hit CPU cost	631.8 ns/op	666.5 ns/op	+5.5% (stable)
Labels cache-hit memory	48.0 B/op	48.0 B/op	0.0% (stable)
Labels cache-hit allocations	3.0 allocs/op	3.0 allocs/op	0.0% (stable)
Labels cache-bypass CPU cost	773.2 ns/op	779.0 ns/op	+0.8% (stable)
Labels cache-bypass memory	53.0 B/op	52.0 B/op	-1.9% (stable)
Labels cache-bypass allocations	3.0 allocs/op	3.0 allocs/op	0.0% (stable)

State

• Coverage, compatibility, and sampled performance are reported here from the same PR workflow.
• This is a delta report, not a release gate by itself. Required checks still decide merge safety.
• Performance is a smoke comparison, not a full benchmark lab run.
• Delta states use the same noise guards as the quality gate (percent + absolute + low-baseline checks), so report labels match merge-gate behavior.