flag_cluster: fix the three findings, and pin the mordant pack by rev

[alii]

language CI went red with no commit in this repo. Cargo.toml pinned the mordant pack by git URL with no rev, so it tracked mordant master — and mordant#4 put flag_cluster there.

Three findings, three different right answers. -D warnings aborts dylint at the first failing crate, so the first was masking the other two: neither had ever been seen until it was fixed.

The threshold is 3 bools, not 2

flag_cluster.rs reads config.flag_cluster_min_bools.max(2), and MordantConfig defaults it to 3; the .max(2) floor only applies to a workspace with no dylint.toml at all. language has one.

That is load-bearing: a two-of-four collapse on Compiler would have cleared the lint without fixing anything, and it is why FmtArgs is fixed by collapsing three flags and leaving one.

vm::http::HeadFlags — 8 representable, 8 observable, 6 distinct

Connection is a token list and may be repeated, so close, keep-alive is a real wire state — but (close, keep_alive) and (close, ¬keep_alive) behave identically everywhere, so only 3 of the 4 are distinct.

Modelled the observation, named: enum ConnTokens { Neither, Close, KeepAlive, Both }, with accumulation across repeated fields an explicit union rather than two independent |=. Intent-modelling gives 6 states but changes the record's arity, needs three new AbiSlots behind a new public Persistence type in h1.scrl, and reverses a decision the code states twice ("raw findings, not decisions") — a stdlib API change riding on a CI-unbreak.

The RFC pointer in the ticket was off by a subsection: §9.1 is Establishment and says nothing about tokens; the opinion is §9.3 Persistence, which h1.scrl already cited correctly at should_close.

bytecode::compiler::Compiler — 2^4 = 16 → 16, argued suppression

No state dropped, because none is illegal. Every write site walked: check_only is construction-only, collect_hover_facts is a post-construction latch, retain_namespaces is scoped by with_retained_namespaces, walking_module_statements is mem::replace-scoped. No pair is exclusive.

And the all-true corner has a witness the ticket lacked: IncrementalSession::new_from_source → new_compiler(None, true), sets collect_hover_facts, then register_prelude runs under with_retained_namespaces, inside which analyse_module sets walking_module_statements. All four at once. All-false is an ordinary compile.

The argument lives in the code above the attribute, not in a baseline entry, so MORDANT_BASELINE_WRITE=1 cannot eat it. Supporting evidence: this struct's mode bools that did carry an invariant are already enums (UnusedBindings, ModuleScope) — the file's owners drew this line already, and these four sit on the other side of it.

FmtArgs — 2^4 = 16 → 8, and the lint was pointing at a live defect

--stdin --stdout parsed, and cmd_fmt's if args.stdin branch returned before ever reading args.stdout — the flag was silently ignored. path was excluded from --stdin by a clap attribute, not by the type.

clap's conflicts_with already excluded 6 of 16 at parse time; of the remaining 10, (stdout, stdin) in both debug settings behaved identically to stdin alone. So 8 were meaningfully distinct — 4 output modes × 2 debug — and the type now represents exactly those.

FmtTarget is Stdin or Files { path, action }, so path lives only in the arm where it is legal. derive(Args) cannot produce this, so Args/FromArgMatches are hand-written — the Command they build is what --help and clap_mangen render, and every flag keeps its text.

Behaviour change worth your judgement: --stdin --stdout now exits 2 with the argument '--stdin' cannot be used with '--stdout' rather than being accepted and ignored. Previously accepted; never did anything.

The pin

rev = "eba57c33...", verified it took — the run's DYLINT_METADATA names that rev and the checkout appeared. Carried over from Scarlet's pin block, including that rm -rf target/dylint is what makes a bump take. Deliberately not carried: Scarlet's sentence about xtask/src/check/mordant.rs catching that staleness — language has no such gate, so it is a manual step here.

The pin is now something somebody must move. t195-on-master (264bc0d) is already waiting on the other side of it.

Also required, and measured rather than predicted

scarlet_core gains [lints.rust] for unexpected_cfgs: cargo dylint sets --cfg dylint_lib="mordant" and plain cargo does not, so the cfg_attr made clippy -- -D warnings fail. A [workspace.lints] table would scale better but touches every manifest.

Plants

Suppression removed → rc=1 naming Compiler at mod.rs:411. main.rs reverted → rc=1 naming FmtArgs at main.rs:90. Old struct literal → E0560, four fields. Old HeadFlags field spelling → E0609, "available fields are: conn, expect_100_continue".

Every dylint run with rm -rf target/dylint, a touch, and a .rc deleted immediately before — one earlier read on this branch showed stale rc=0 files from a previous run and was discarded (T-83).

Gates

fmt 0 · clippy -D warnings 0 · test --workspace 0 (43 binaries / 1303 passed) · gen-editor-syntax 0 · SCARLET_GC_STRESS=1 0 · cargo dylint --all 0.

No suppression in the tree lacks an argument. The only one is Compiler's, and its argument is the eleven lines above it.