fix(stdlib): #4917 — real semantics (or loud warn) for silently-ignored adapter options and no-ops

[proggeramlug]

Closes #4917 (stub-elimination epic #4919). Code-only per contributor convention — no version bump / changelog; maintainer folds metadata at merge.

Per-checkbox status, in the issue's order:

Options accepted but ignored

• zlib stream factories — options.level is now honored by every deflate-family factory (createGzip/Deflate/DeflateRaw/...), threaded through make_codec_state; .reset() rebuilds the codec at the stream's level. deflateRawSync gains its missing options arg (stdlib fn + codegen native-table arity + FFI decl + dynamic-dispatch arm). A supplied dictionary now warns once instead of silently mis-compressing — the warn lives in the shared js_zlib_validate_options (perry-runtime), so the ext-zlib auto-optimize path inherits it. Brotli/zstd factories warn once when an options object is passed (params shape still unwired). Remaining honest gaps (strategy/memLevel validated-but-unapplied, brotli/zstd params) keep narrowed manifest notes.
• http.Agent — partly stale: on the default ext-http path, per-agent reqwest::Client pooling already honors keepAlive/maxFreeSockets/keepAliveMsecs, and destroy() already drops the cached client (real pool teardown) — un-flagged in the manifest. The true remaining no-ops, keepSocketAlive/reuseSocket, now warn once on both the ext and stdlib paths.
• exponential-backoff — full option surface honored: numOfAttempts, startingDelay, timeMultiple, maxDelay, delayFirstAttempt, jitter: 'full', and the retry(e, attemptNumber) predicate. Bigger find: the adapter never retried async tasks at all (the first Promise was passed through verbatim). Rewritten as a GC-rooted state machine that attaches reactions via js_promise_then and schedules delays through the timer queue — retry-on-rejection with no blocking thread::sleep on the main thread. The old code also read options from the wrong register (declared *const ObjectHeader for an NA_F64 arg).

No-ops returning success

• fastify upgrade — registering an upgrade handler now warns once at on() time (the stored-callback silence was the lie); upgrade requests keep the loud 501 until FastifyInstance.server returns a number — blocks ws/http2 upgrade piggybacking #1113 lands real dispatch.
• worker.ref()/unref() — already real: the cited worker_threads.rs:452,638,643 are MessagePort no-ops, not Worker. Worker.ref/unref flip WorkerRecord.refed, which gates js_worker_threads_has_pending. Verified behaviorally both directions (unref'd live worker → process exits in 0.46s; refed worker → holds until terminate()). Stale manifest stub notes dropped.

Wrong-shaped results

• mongodb findOne() — resolves a parsed document object (doc.field works). The blocking JSON.parse(...).foo → NaN bug noted in the code comment is fixed on main (probed). BSON-specific types surface in relaxed extended-JSON shape (_id.$oid). Parse happens in the main-thread converter (no worker-arena hazard).
• mysql2 column metadata — field.type/columnType now carry the numeric MySQL wire type ID via a name→ID map (sqlx 0.8 keeps the raw ColumnType byte pub(crate); the name strings are a bijection over type+flags). length stays 0 and is documented — max_size is not reachable through sqlx's public API. Both perry-stdlib/mysql2 and perry-ext-mysql2 updated.
• pg column metadata — dataTypeID is the real numeric type OID (PgTypeInfo::oid()), tableID/columnID come from the RowDescription via relation_id()/relation_attribute_no() (0 for expression columns, like Node), plus dataTypeSize/dataTypeModifier (-1 sentinel) and format: "text" for node-pg shape compat; field objects now carry a keys array so by-name access works. Both stdlib and ext twins updated.

Manifest / docs

• #4917 stub inventory 18 → 9 (remaining: 3 compressor factories partial-strategy + 4 brotli/zstd params + 2 Agent socket hooks), with per-entry notes reworded to describe the remaining divergence. stub_inventory.rs keystone spot-checks updated (findOne/backOff/ref/unref/destroy moved to must-NOT-be-stub). perry.d.ts / reference.md regenerated via scripts/regen_api_docs.sh.

Verification

• New test-files/test_gap_zlib_4917_level.ts: byte-for-byte PARITY-MATCH vs node v26 (level honored on streams + deflateRawSync, RangeError on level 99, round-trips).
• exponential-backoff e2e: retry predicate called with attempt numbers, success after N rejections, retry: false stops after 1 call, numOfAttempts: 3 exhausts after exactly 3 calls rejecting with the last error, delays measured.
• Worker ref/unref e2e both directions (above).
• dictionary warn-once verified on stderr.
• 0-regress: test_parity_zlib, test_gap_zlib_fs_assert_2935_2752_2971, test_parity_worker_threads MATCH vs node; test_gap_zlib_3285_params diff is pre-existing (byte-identical output from a pre-change baseline binary — the ext-zlib .params() flip isn't active in this environment).
• cargo test -p perry-api-manifest (36 green incl. stub inventory drift guard), ext-http/-mysql2/-pg unit tests (28 green), full release build of all touched crates green, file-size gate green.

[proggeramlug]

Rebased onto current main and added a second commit unbreaking the lint job: #4994 (da4f8be) was admin-merged with lint red, so every PR merge-ref inherited two audit failures from it — a GC_STORE_AUDIT marker one line out of window in array/sort.rs, and two un-allowlisted GcHeader probes (array/from_concat.rs, object/prototype_chain.rs). All four lint gates (fmt, file-size, GC store-site, addr-class) verified green locally on the rebased tree.