noyalib-debugging-playbook

name	noyalib-debugging-playbook
description	Symptom-driven triage for noyalib (pure-Rust YAML 1.2). Load when a test fails, a parse result surprises you, spans point at the wrong bytes, budget errors fire unexpectedly, `from_str::<Value>` and `from_str::<MyStruct>` disagree, `no_std` breaks despite green CI (symptom triage here; the cache-poisoning CI doctrine that PREVENTS it lives in noyalib-ci-and-release), column desync / spurious indent errors show up, or you need to answer "why does the parse differ?" quickly. Runbook voice with the rationale embedded — for a mid-level engineer or Sonnet-class model with zero project context.
date	"2026-07-07T00:00:00.000Z"

noyalib debugging playbook

Repo: github.com/sebastienrousseau/noyalib — pure-Rust YAML 1.2 library, workspace crates/noyalib, current release v0.0.14. Run every command below from the repo root.

Reference material (full 6-row loader-routing table, full 14-row symptom table, full 12-knob/6-variant budget inventory, and traps (d) NaN key equality + (e) cargo vet fmt strips TOML comments): see reference.md in this directory. Load when the top-N runbook paths below don't pin the bug, when you need the exhaustive symptom row set, or when you need the complete BudgetBreach-vs-other-Error mapping.

THE cardinal triage question

Which of the THREE loaders is my bug in?

from_str_with_config in crates/noyalib/src/de.rs routes to one of three code paths. Almost every "same input, different result" report resolves once you know which path fired.

The routing decision lives around de.rs:396-499. Read it — it changes. As of v0.0.14 the shape is:

// crates/noyalib/src/de.rs (approx line 415)
let value_target_bypass = is_value_target::<T>() && config.tag_registry.is_none();
let stream_eligible = config.merge_key_policy == MergeKeyPolicy::Auto
    && !config.ignore_binary_tag_for_string
    && config.policies.is_empty()
    && properties_inactive(config)
    && includes_inactive(config)
    && !value_target_bypass;

Routing table (top 3 rows — full 6-row matrix in reference.md §routing-full)

T	Config knobs	Path	Where
any typed T (not Value)	all defaults	streaming (streaming.rs)	de.rs stream_eligible == true
Value	tag_registry.is_none() (i.e. no registry active)	NoSpanLoader (span-free)	de.rs:454 is_value_target::<T>() fast path
any typed T	any non-default config knob above	Loader (span-full) then serde	de.rs:472 #[cfg(feature = "std")] block

Nuance you will hit: v0.0.14 excludes Value-target from streaming unless a TagRegistry is active. The NoSpanLoader owns the distinct-typed KeyCollision guard and the DoS budgets. See bc8f798 fix(loader): plug distinct-typed key collision on Value fast path. Rationale + Spanned<T> / Value+registry / CST rows: see reference.md §routing-full.

The discriminating experiment

Feed the same input through all three loaders and diff:

let v_value: Result<Value, _>       = noyalib::from_str::<Value>(src);
let v_typed: Result<MyStruct, _>    = noyalib::from_str::<MyStruct>(src);
let v_cst:   Result<Document, _>    = noyalib::cst::parse_document(src);

Divergence localises the bug:

• Value OK, typed Err → streaming deserializer (streaming.rs).
• Value Err, CST OK → NoSpanLoader vs green-tree walker divergence; dup_key_spans.rs is the canonical parity suite.
• Value OK, CST Err → span-full path or CST-specific walk.
• All three disagree on numbers/bools → scalar resolution split across three sites; check 1225-1367 in parser/loader.rs.

This is exactly the cross-check fuzz/fuzz_targets/fuzz_no_span_loader.rs runs; it exists because the silent-collapse bug survived 4000+ green tests when no test compared Value and CST paths directly.

Symptom table

The top 5 rows by hit rate are below; the full 14-row table with every variant (span-fix rows for f8fe929 / 1f76951 / af7a9bd / 1d72687, no_std cache poisoning, scanner CR/BOM row, budget-error wart row, next_value_seed misuse, RecursionLimitExceeded on shallow-but-wide) is in reference.md §symptom-full. Each row cites the file/test/commit that pins the behaviour so you can verify against source, not against my memory.

Symptom	Likely cause	First check	Verify against
"Parses but data missing (some keys gone)"	DuplicateKeyPolicy collapsing genuine duplicates, or (pre-0.0.14) distinct-typed collision silently overwriting	Set config.duplicate_key_policy = DuplicateKeyPolicy::Error and re-run	tests/dup_key_spans.rs::distinct_typed_keys_collide_loudly (crates/noyalib/tests/dup_key_spans.rs:186)
"Works as struct T, fails as Value" (or vice versa)	Loader divergence — streaming vs NoSpanLoader vs Loader	Run the discriminating experiment above	tests/no_span_loader_parity.rs (crates/noyalib/tests/no_span_loader_parity.rs)
"Span points at wrong bytes"	CST walker; five separate fixes landed in v0.0.14 cycle	git log --oneline -20 -- crates/noyalib/src/cst/	Commits f8fe929, 1f76951, af7a9bd, 1d72687
Span includes trailing blanks	span_at trims trailing blanks; raw event spans don't	Compare doc.span_at(path) vs Spanned<T> extents	tests/dup_key_spans.rs::spanned_fields_stay_aligned_after_duplicate_key
"Budget error unexpectedly"	One of the 12 ParserConfig budget knobs fired (only 6 flow into BudgetBreach; others surface as RecursionLimitExceeded, RepetitionLimitExceeded, Serialize, or Parse — see budget summary below)	Read the error message + variant	crates/noyalib/src/error.rs:161 BudgetBreach enum

Budget inventory (summary)

Load-bearing counts: ParserConfig has 12 budget knobs; only 6 of them flow through Error::Budget(BudgetBreach::...) (MaxEvents, MaxNodes, MaxTotalScalarBytes, MaxDocuments, MaxMergeKeys, AliasAnchorRatio) at crates/noyalib/src/error.rs:161. The other 6 surface elsewhere:

• max_depth → Error::RecursionLimitExceeded
• max_alias_expansions → Error::RepetitionLimitExceeded
• max_mapping_keys / max_sequence_length → Error::Serialize (spelling wart — callers pattern-match on msg.contains(...); do not "fix" without checking callers)
• max_document_length → Error::Parse
• max_include_depth → include-feature specific

Full knob-by-knob mapping, BudgetBreach variant field shapes, the exact wart-message strings, and the test lines that assert them: see reference.md §budget-full.

Traps with stories

(a) cargo fuzz needs an explicit target on Apple Silicon

Trap: Running cargo +nightly fuzz run fuzz_no_span_loader on an M-series Mac errors with E0463: can't find crate for 'std'.

Why it bites: cargo fuzz defaults to x86_64-apple-darwin; recent nightlies don't ship that std by default on Apple Silicon. The error mentions std, not target, so it looks like a toolchain bug.

Avoid: always pass --target aarch64-apple-darwin on Apple Silicon:

cargo +nightly fuzz run fuzz_no_span_loader --target aarch64-apple-darwin -- -max_total_time=60

(b) Tests passing does not mean correct — cross-loader parity is the load-bearing check

Trap: the distinct-typed key silent-collapse (1: a then "1": b overwrites) survived 4000+ green tests.

Why it bites: every test looked at one loader in isolation. The Value-typed tests exercised NoSpanLoader; the struct-typed tests exercised streaming; the span tests exercised Loader. None compared outputs across paths, so an entry silently vanishing on one path but being caught on another produced 100% green.

Avoid: any behavioural fix that touches key handling, budgets, or collision must add a parity test to tests/no_span_loader_parity.rs AND update fuzz/fuzz_targets/fuzz_no_span_loader.rs (which cross-checks from_str::<Value> vs cst::parse_document).

(c) cwd left in fuzz/ breaks cargo — "not a member of the workspace"

Trap: After a fuzz session you run cargo test -p noyalib and get error: package ID specification 'noyalib' did not match any packages or current package believes it's in a workspace when it's not.

Why it bites: fuzz/Cargo.toml is its own crate outside the workspace. If your shell is still in fuzz/, cargo picks up that manifest instead of the workspace root.

Avoid: cd "$(git rev-parse --show-toplevel)" at the top of any multi-step script. Every noyalib maintenance script (.git/*.sh) starts with that line for exactly this reason.

Two more traps — (d) NaN key equality is a landmine for the collision check (why we don't use HashSet<Value> for key dedup), and (e) cargo vet fmt strips TOML comments so exemption rationale must go in the commit — are in reference.md §traps-extended.

Discriminating experiments — copy-paste commands

Always run from the workspace root:

cd "$(git rev-parse --show-toplevel)"

Minimal reproducer template

// tests/scratch_repro.rs   (put in crates/noyalib/tests/)
use noyalib::{Value, from_str, cst::parse_document};

#[test]
fn repro() {
    let src = "PUT_YAML_HERE\n";
    let v  = from_str::<Value>(src);
    let d  = parse_document(src);
    eprintln!("value = {v:#?}");
    eprintln!("cst   = {:?}", d.as_ref().map(|d| d.to_string()));
}

Run one focused test with backtrace and stdout:

RUST_BACKTRACE=1 cargo test -p noyalib --test dup_key_spans -- --nocapture
RUST_BACKTRACE=1 cargo test -p noyalib --test no_span_loader_parity distinct_typed_key_collision -- --nocapture
RUST_BACKTRACE=1 cargo test -p noyalib --test issue_46 no_span_loader_honours_max_depth -- --nocapture

Run one fuzz target for 60s (Apple Silicon):

cargo +nightly fuzz run fuzz_no_span_loader \
  --target aarch64-apple-darwin -- -max_total_time=60

For x86_64 hosts drop --target.

Inspect the routing gate (re-verify after any de.rs change):

grep -n "stream_eligible\|is_value_target\|value_target_bypass" crates/noyalib/src/de.rs

Cross-loader smoke check on a single input, one-shot:

cargo test -p noyalib --test no_span_loader_parity -- --nocapture

CST span probe from a REPL-ish test:

let doc = noyalib::cst::parse_document(src).unwrap();
if let Some((s, e)) = doc.span_at("some.path") {
    eprintln!("[{s}..{e}] = {:?}", &doc.source()[s..e]);
} else {
    eprintln!("no span for path");
}

When NOT to use this skill

• For the full history of settled investigations (why a fix is shaped the way it is, what earlier attempts failed) → noyalib-failure-archaeology.
• For measuring instead of guessing (benchmarks, allocation profiles, flamegraphs) → noyalib-diagnostics-and-tooling.
• For release / version-bump mechanics → noyalib-change-control.

This skill is scoped to "what broke, where do I look, what experiment discriminates?" — not history, not measurement, not release choreography.

Provenance and maintenance

This provenance section covers both SKILL.md and reference.md in this directory — one truth source for the split.

Ground-truth files consulted (verify each still says what this skill claims when you update it):

• crates/noyalib/src/de.rs — routing gate stream_eligible (~L415), is_value_target (L170), Value-target fast path (L454).
• crates/noyalib/src/parser/loader.rs — Loader struct (L241), NoSpanLoader struct (L861), NoSpanFrame::typed_keys parallel vec (L843), budget spellings (L575, L675, L1063, L1135).
• crates/noyalib/src/streaming.rs — raw_str_mode in next_key_seed (L1346-1348), finished guard on next_value_seed (L1363).
• crates/noyalib/src/cst/document.rs — span_at (L193), resolve_path_in_green + typed-cache fallback (L201-L211), trim_value_span behaviour.
• crates/noyalib/src/error.rs — BudgetBreach enum (L161), Error::Budget variant (L542).
• crates/noyalib/tests/no_span_loader_parity.rs — parity contract.
• crates/noyalib/tests/dup_key_spans.rs — span/typed-view alignment on duplicates + distinct-typed collision refusal.
• crates/noyalib/tests/issue_46.rs — NoSpanLoader depth-limit regression + RecursionLimitExceeded shape.
• fuzz/fuzz_targets/fuzz_no_span_loader.rs — cross-loader divergence cross-check.
• Commits: bc8f798, 0fff9dd (distinct-typed collision), f8fe929, 1f76951, af7a9bd, 1d72687 (five CST span fixes), 70c77cb (lone-CR scanner fix), 3304f4c (BOM v0.0.10).

Re-verify the routing table before trusting it — one command:

grep -n "stream_eligible" crates/noyalib/src/de.rs

If the surrounding boolean expression grew a new conjunct, update the routing table above. This is the single most load-bearing claim in the skill.

Last updated: 2026-07-08 against feat/v0.0.14 (HEAD == 5c41ef3 chore(supply-chain): update cargo-vet exemptions). Split into SKILL.md + reference.md on 2026-07-08.