PromQL parser-fork gaps (umbrella) — 72% of corpus rejections, all in promql-parser #108

Umbrella — PromQL parser-fork gaps (tracking)

233 of the ~322 PromQL corpus rejections (72%) are L1 parse failures in the private promql-parser fork (ProjectASAP/promql-parser, branch asap), before the lowerer ever runs. This repo consumes the parser via a git pin; the fixes land in the fork, and the only change here is bumping that pin once each lands.

Correction (2026-07-10): none of these are "mirror upstream"

An earlier read of this issue assumed some gaps (smoothed/anchored/start/end) were just a rebase onto GreptimeTeam upstream away. That was checked against the actual grammar and it does not hold — GreptimeTeam upstream lags Prometheus (Go) on every one of these:

• smoothed / anchored are stub tokens even upstream — src/parser/promql.y:124 marks them %expect-unused (declared, no production consumes them) on both the fork and GreptimeTeam main.
• trim operators, duration expressions, and callable start()/end() have no token / AST anywhere upstream.

The fork's asap commit (c51beaf) added the experimental functions (info, histogram_quantiles, min_of/max_of/step/range, the *_over_time reducers) but explicitly deferred the grammar/lexer features. So all five gaps are net-new parser work on the asap branch; there is nothing to mirror.

Split (all in ProjectASAP/promql-parser)

Ordered easiest-first: #5 (bug fix) → #1 (token exists) → #4 → #3 → #2 (biggest).

The remaining ~89 non-parser rejections are lowering gaps (tracked elsewhere) and deliberate rejections (out-of-range φ, __name__ regex per #67).