Filter param-env caller bounds to the goal-relevant subset in canonical queries

[xmakro]

Every canonical query (the type ops, evaluate_obligation, query normalization, implied_outlives_bounds) keys its cache on the full ParamEnv. Two functions proving the literally identical goal get distinct cache keys whenever their where-clauses differ, even when none of those clauses can possibly influence the goal. The degenerate case is a trait method with where Self: 'static: every impl gets a distinct caller bound (S0: 'static, S1: 'static, ...), so the same MIR typeck goals re-run once per impl.

This adds a relevance filter at a new canonicalize_query_with_env_filter entry point, used by the trait-solving query family. For param-free goals it reduces the env to the subset of caller bounds that could participate in proving the goal:

• Trait-like clauses (trait, projection, host-effect, const-arg/const-evaluatable) are kept only if their head is param-free. Generic params are rigid during solving, and a subgoal of a param-free goal can only mention params introduced by the right-hand sides of already-kept clauses, which a bitmask fixpoint accounts for (e.g. a kept <Vec<u8> as X>::Out == T pulls in T's clauses).
• Type-outlives clauses are kept for alias or region-carrying subjects; a fully concrete rigid subject (S0: 'static) is vacuous inside a canonical query, since queries only collect region constraints and the env's outlives clauses are consulted caller-side against the unfiltered env.
• Region-outlives, well-formed, and unstable-feature clauses are always kept.

Goals mentioning inference vars, non-region bound vars (embedded canonical values, e.g. user type annotations), or generic params keep the full env. The gating to param-free goals keeps the per-call cost at a few flag checks plus one hash lookup on a (caller_bounds, mask) cache; there is no goal walk on the hot path. Method probing intentionally stays on the unfiltered canonicalize_query, since it manufactures inference vars that are matched against the env beyond what the goal itself requires.

Clean A/B on the full local rustc-perf suite (from-scratch ThinLTO+jemalloc stage2, instructions:u, vs the 2026-08-09 base): mean -0.311% over 551 cells, 191 significant wins, 0 regressions. Highlights:

benchmark	change
wg-grammar-new-solver check/debug/opt full	-2.64% / -2.54% / -2.32%
hyper-1.6.0 check full / incr-full	-2.25% / -2.38%
image-0.25.6 check full / incr-full	-2.20% / -2.29%
projection-caching check full	-1.80%
diesel-2.2.10 (all profiles, full + incr-full)	-1.08% .. -1.21%
serde-1.0.219 (all profiles)	-0.87% .. -0.94%
syn-2.0.101 check full	-0.88%
bitmaps-3.2.1	neutral

Microbenchmark and validation detail (stage1, same-binary on/off via RUSTC_ENV_FILTER_DISABLE):

• Stress case (400 impls with where Self: 'static, identical concrete method bodies): 3.034G -> 2.650G instructions, -12.65%; the control without the where-clause is unchanged, i.e. the filter recovers about two thirds of the pure fragmentation excess.
• Filter stats on diesel (RUSTC_ENV_FILTER_STATS): 377k canonical queries, 90% served from the filtered-env cache, ~8k distinct filtered envs; executions drop across the family (evaluate_obligation -25%, type_op_prove_predicate -20%, dropck_outlives hits 5x) and the dep graph loses 28k nodes on this crate.
• ui suite: 21720 passed, 0 failed.

Follow-ups worth exploring: extending filtering to goals that mention params (needs a cheap per-goal param mask, e.g. collected during canonicalization itself; a full-relevance prototype deduped substantially more but the per-call goal walk cost more than it saved), applying the same relevance test to the global selection/evaluation cache keys, and dropping region-outlives clauses once their in-query uses are audited.