v3: fix self-host OOM and five checker/cgen regressions blocking -building-v

[medvednikov]

`./v3 -nocache -building-v -o v4 -no-memory-limit v3.v` was OOM-killed by macOS right after the parse phase (the process allocated at ~20 GB/s in the parallel checker until jetsam sent SIGKILL — `-no-memory-limit` only disables the internal watchdog, not the OS kill). Fixing the OOM then exposed five more self-host blockers. With this PR the full bootstrap chain is green again: v1 → v3(prod) → v4 → v5, ~1.8 s total, ~740 MB steady footprint.

Memory explosion (checker)

• `source_enclosing_fn_has_generic_param` copied the whole source file twice per call (`source[..offset]`, `source[fn_start..]` — ~5 MB per call on checker.v) and was reached for every identifier in `resolve_type`/`check_ident` because the cheap `is_bare_generic_param` guard was applied after the expensive `||` chain. Now index-scans without copies, and both call sites short-circuit on the guard first (behavior-preserving: the chain's result was only ever consumed under that same guard).
• `closest_identifier_span` copied the file suffix once per occurrence of the searched name (O(occurrences × file size) bytes). Now uses `index_after_` and stops after the first word match past the anchor (later matches are strictly farther).

Checker fixes

• `s[a..n] or { ... }` (gated range slice) was typed as the element type `u8` instead of the slice type; `or_expr_payload_type` now recognizes `value == 'range'` index sources (broke `v/util/version.v`).
• `name.str` on a string in infix position resolved to the `str()` method value (`fn () string`) instead of the `&u8` field; declared builtin-struct fields now take priority over builtin method values in `selector_type`.

Transform fixes

• `a, b := match e { ... else { return } }`: the multi-value match lowering bailed out on branches whose tail exits (return/break/continue/noreturn), falling into a broken `_ifexpr` fallback. Exiting branches now keep their body unchanged with no assignment.
• Value-producing if-guards with multi-value destructuring (`x := if a, b, _ := f() { ... }`) bound the whole tuple to the first name; `build_if_value_guard_chain` now emits per-name `.value.argN` bindings like the statement path.

Cgen fixes (-building-v)

• The fixed-storage-const optimization miscompiled `in const_array`: the lowered `array_get(const, i)` shares its const AST node with a `.len` selector base, which the safety scan had whitelisted, so the const was wrongly promoted to a C fixed array. Const references in plain call arguments now always count as dynamic uses.
• `v.token` newly entered the v3.v closure (checker.v imports `v.util`), and `-building-v` template erasure left `scanner_matcher = new_keywords_matcher_trie(...)` calling a deleted function (the keep-list preserves the template AST, but cgen never emits open generic templates). Const initializers that call erased templates are now dropped with their storage; the compiler never reads them, and any new use fails loudly at C compile time instead.

Each fix was reduced to a minimal repro first (match-tuple exit branches, guard destructuring, slice-or, `.str` infix, `import v.token` under `-building-v`), and all repros compile and run correctly after the fix.

[medvednikov]

Rebased on master. Note: the rebase surfaced a separate master regression from 79e0c1c — permanently arming qualify_name_cache on the main checker segfaults every -prealloc build (scoped driver stages grow the cache map inside disposable scope arenas; freeze_interface_impl_names later reads freed buckets in map_get_check). Fixed here as a second commit by restoring the nil-unless-armed discipline; parallel cgen forks keep their private armed instances. Verified: v1 → v3 → v4 → v5 bootstrap chain green with -gc none -prealloc.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3d80692d60

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Add regression coverage for gated range slicing

This changes Option/Result payload typing for s[a..n] or { ... }, including distinct string, dynamic-array, and fixed-array paths, but the commit adds no test file. Without a focused regression test, the original element-vs-slice typing failure can recur unnoticed; please add coverage for these gated range cases as required for Option/Result behavior changes.

AGENTS.md reference: AGENTS.md:L657-L659

Useful? React with 👍 / 👎.