suggest-native-expression

name	suggest-native-expression
description	Use when picking the next Spark expression to implement natively in Comet, or when asked whether a specific codegen-dispatched expression is worth a native Rust implementation. Scores the candidate on how likely a native path can be 100% Spark-compatible and how much throughput or allocation it would save versus the JVM codegen dispatcher, records the verdict in the per-expression audit log so disqualified candidates are not re-litigated, and files a GitHub issue for the recommendation.
argument-hint	["expression-name"]

Assess whether a Spark expression that Comet runs through the JVM codegen dispatcher is worth implementing natively in Rust, and propose the best candidate as a GitHub issue.

This skill does not implement anything. Its deliverables are exactly three:

1. A verdict on two axes: compatibility confidence (can a native path be 100% Spark-compatible?) and native upside (how much throughput or allocation does the native path save versus codegen dispatch?).
2. A dated Native candidate line in the expression's page under docs/source/contributor-guide/expression-audits/, so a future run can see that this expression was already assessed and why it was or was not taken.
3. One filed GitHub issue for the recommended (or deferred-with-a-named-blocker) candidate.

$ARGUMENTS may name a single expression to assess. If it is empty, run in survey mode (Step 1) and pick the best candidate from the whole pool.

Background reading

Read these before scoring anything. Do not score from memory: the routing model and the cost of the dispatcher are both easy to misremember, and every judgment in this skill depends on them.

• docs/source/contributor-guide/roadmap.md, section "Native Coverage for Codegen-Dispatched Expressions". This is the project-level statement of why this work matters and it is the section the filed issue should reference.
• docs/source/user-guide/latest/compatibility/index.md, section "Native and codegen-dispatch implementations". Defines the two paths and which one is the default.
• docs/source/user-guide/latest/expressions.md, the Implementation legend and tables. Useful for a candidate's SQL name, category, and status, but see Step 1: this table is not the pool.
• docs/source/contributor-guide/optimizing_expressions.md, the "Proven techniques" table. If no technique in that table applies to the candidate, the native upside claim is weak.
• spark/src/main/scala/org/apache/comet/serde/CometExpressionSerde.scala for the CometCodegenDispatch / CodegenDispatchFallback / NativeOptInAvailable contracts.
• spark/src/main/scala/org/apache/comet/codegen/CometBatchKernelCodegen.scala for what the dispatcher can and cannot compile (isSupportedDataType, canHandle).

What the dispatcher actually costs

Getting this right is the whole point of the skill. Overstating the dispatcher's cost produces issues that waste a contributor's week on a 2% win.

What codegen dispatch does cost:

• One JNI round trip per batch, out of native execution and back.
• Per-row evaluation of Spark's generated Java. No vectorization, no SIMD, no dictionary awareness, no bulk buffer operations.
• Per-row JVM heap allocation for every non-primitive intermediate and result: UTF8String, Decimal/BigDecimal, GenericArrayData, ArrayBasedMapData, boxed values, plus whatever Spark's own doGenCode allocates (formatters, StringBuilder, temporary arrays).
• That allocation lands on the JVM heap, outside Comet's native memory pool, so it is invisible to Comet's memory accounting and it feeds GC pressure that shows up as task-level jitter rather than as an expression cost.
• Per-row reads of nested Arrow data go through CometArrayData / CometMapData / CometSpecializedGettersDispatch wrappers, which allocate per row per nesting level.

What codegen dispatch does not cost, and therefore is not available as an argument for going native:

• It does not convert the whole batch to rows and back. The kernel reads Arrow vectors directly.
• String reads are zero-copy (UTF8String.fromAddress).
• The Janino kernel is compiled once per (expression, schema) pair, not per batch.
• A NullIntolerant root short-circuits null rows without evaluating the body.
• It is byte-exact with Spark by construction, so it needs no compatibility work at all. A native replacement starts with a compatibility debt that the dispatcher does not have.

The corollary: a native implementation is worth proposing when the per-row JVM work is allocation-heavy or vectorizable, not merely because "native is faster than JVM".

Step 1: Choose the candidate

If $ARGUMENTS names an expression, skip to Step 2 but still run the exclusion checks below against it.

Build the pool

The serde definitions are the source of truth, not the docs table. Start there:

# Sub-pool A: expressions whose only path is the dispatcher, plus those with a native opt-in.
grep -rhoE "CometCodegenDispatch\[[A-Za-z0-9]+\]" spark/src/main/scala/org/apache/comet/serde/*.scala |
  sed 's/.*\[//;s/\]//' | sort -u
# Which of those also have a native path (rendered "Hybrid" in the docs)?
grep -rn "NativeOptInAvailable\|CodegenDispatchFallback" spark/src/main/scala/org/apache/comet/serde/*.scala

Then cross-reference docs/source/user-guide/latest/expressions.md for each candidate's SQL name, category, and Implementation label. Do not build the pool from that table alone: its Implementation column is derived from a function-registry lookup, so a dispatcher-routed serde whose SQL name does not resolve is rendered — and disappears from a table-driven grep. mask is the worked example: CometMask extends CometCodegenDispatch[Mask], but its row reads —. At the time of writing the counts are 67 classes extending CometCodegenDispatch against 58 rows labelled Codegen dispatch, so a table-driven pool silently loses candidates.

Two sub-pools, both in scope, with different proposed approaches:

• A. Dispatch-only (a plain CometCodegenDispatch, no native path). The proposal is a new native kernel or an upstream wiring.
• B. Hybrid (NativeOptInAvailable / CodegenDispatchFallback: a native path exists but is Incompatible, so the dispatcher is still what runs by default). These are in scope because the user-visible default is identical to sub-pool A, and so is the win from fixing it. The proposal is different: close the compatibility gap, or add a runtime guard that routes only the genuinely incompatible inputs back through the dispatcher, so the native path can become the default. The mechanics belong to audit-comet-expression; this skill's job is to establish that the win is worth someone's time. Roughly 42 expressions sit here, so skipping them would drop the largest block of candidates on a technicality.

Out of scope: rows marked — or 🔜 Planned where Comet has no path at all and the whole operator falls back to Spark. Those are implement-comet-expression territory, because the question is "implement it at all" rather than "is native better than dispatch". Their upside is genuinely larger than anything in this pool, so if you are looking for the biggest win available, say so and point the user at that skill rather than silently ranking within this pool as if it were the whole space.

Exclude

1. Families the project has ruled out. Read the "Not currently planned" section of docs/source/user-guide/latest/expressions.md and drop anything it covers. Proposing one of these needs a project-direction argument that this skill does not have.

2. Already-assessed expressions. This is the check that makes the skill cumulative:

   grep -rn "Native candidate (assessed" docs/source/contributor-guide/expression-audits/
   

   An expression with a Disqualified line is off the pool unless the named blocker in that line has demonstrably changed (an upstream kernel landed, DataFusion gained the function, the Spark semantics were simplified in a new version). If it has changed, say so explicitly in the new assessment and add a new dated line. Never delete or rewrite an old line.

3. Expressions with an existing issue. Search open and closed:

   gh issue list --repo apache/datafusion-comet --search "<function> native in:title,body" --state all --limit 10
   

   If a candidate match comes back, open it (gh issue view <N> --repo apache/datafusion-comet) and confirm it really covers a native implementation of this expression. If it does, that expression is done: pick another rather than filing a duplicate.

Rank what is left

Rank the remaining pool by a cheap first-pass read of the two axes below, then assess the top candidate properly. Prefer expressions that appear in TPC-H or TPC-DS, or in common ETL shapes (string cleanup, date bucketing, JSON field extraction), over expressions that are rare in real queries. To check workload presence:

grep -rli "<function>(" spark/src/test/resources/tpcds/ spark/src/test/resources/tpch/ 2>/dev/null

State the pool size, the exclusions, and why the chosen candidate topped the ranking. A ranking you do not show is a ranking the reader cannot challenge.

Step 2: Establish how the expression runs today

Find the serde and read it:

grep -rn "class\|object" spark/src/main/scala/org/apache/comet/serde/*.scala | grep -i "<Expr>"

Record:

• Which trait it uses, which decides the sub-pool and therefore the shape of the proposal. A plain CometCodegenDispatch[T] means dispatch is the only path (sub-pool A). NativeOptInAvailable or CodegenDispatchFallback means a native path already exists but is off by default (sub-pool B), so the candidate is not "write a kernel" but "make the existing kernel the default": establish which inputs actually diverge, and whether the divergence is detectable well enough to guard. Read the existing getIncompatibleReasons() and treat each listed reason as a claim to verify, not a given.
• Whether the dispatcher can actually compile it, per CometBatchKernelCodegen.canHandle and isSupportedDataType. If some input or output type is outside that surface, those cases fall back to Spark, not to the dispatcher. Whole-operator Spark fallback is a much larger cost than a JNI round trip, and it raises the upside score materially. Say which cases these are.
• Whether the expression is disabled or gated by any config (spark.comet.expression.<Name>.enabled, allowIncompatible, spark.comet.exec.scalaUDF.codegen.enabled).

Step 3: Score compatibility confidence

The question is narrow: can a native Rust implementation match Spark bit-for-bit, for every input, on every supported Spark version? "Close enough" is not a passing answer, because the dispatcher it would replace is exact.

Read the Spark implementation across versions

Derive the version list from the build rather than hardcoding it, so the assessment does not go stale as Comet drops and adds Spark support:

grep -oE "<spark.version>[0-9.]+</spark.version>" pom.xml | sed -E 's/<[^>]+>//g' | sort -u

Read eval / nullSafeEval / doGenCode, inputTypes, dataType, ANSI branches, and any config the expression reads, for each of those versions. The class body is usually short enough to fetch directly instead of cloning:

curl -sSfL "https://raw.githubusercontent.com/apache/spark/v<version>/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/<file>.scala"

Diff the class body between adjacent versions rather than eyeballing it. Note behavior that differs: each divergence is a shim the native path would need, and "identical across all supported versions" is itself a finding worth stating, because it means no shim.

Check upstream first

An existing upstream implementation is the single strongest positive signal, both for compatibility and for effort:

grep -rn "fn name" ~/.cargo/registry/src/*/datafusion-spark-*/src/function/ 2>/dev/null | grep -i "<function>"
grep -rn "fn name" ~/.cargo/registry/src/*/datafusion-functions-*/src/ 2>/dev/null | grep -i "<function>"

If datafusion-spark has it, read the implementation and diff its semantics against Spark's before crediting it. Upstream functions in that crate aim for Spark compatibility but are not all complete. If it matches, the proposal becomes a wiring job (wire-datafusion-function), which is a much cheaper issue to file and a much easier one to pick up.

Hazard checklist

Each hazard that applies is a strike against compatibility confidence. Check every one and report the result, including the ones that do not apply, so the reader can see the checklist was run rather than skimmed.

Hazard	Why it threatens 100% compatibility
JVM formatting APIs (java.text.DecimalFormat, String.format, SimpleDateFormat, DateTimeFormatter) as in format_number, format_string, to_char	Locale, pattern, and rounding behavior would have to be reimplemented, not merely ported
java.util.regex features (backreferences, lookaround, possessive quantifiers, named groups)	The Rust regex crate deliberately omits these, so some patterns cannot be supported at all
Collation (StringTypeWithCollation, CollationSupport.X.exec) on Spark 4.0+	ICU-backed collation is not propagated natively today (#4496)
BigDecimal / MathContext precision and rounding, precision-scale promotion, overflow-to-null	Rounding mode and overflow points are easy to get subtly wrong and produce silent wrong results
Timezone and calendar (JDK tzdb versus chrono-tz, DST edges, hybrid Julian/Gregorian rebase, spark.sql.legacy.* flags)	Divergences appear only on specific dates and zones, so tests rarely catch them
ANSI error class and message parity	An error with the wrong class or message is an observable difference; see the ANSI section of the compatibility guide
Float and double text conversion (Double.toString shortest round-trip), StrictMath versus f64 intrinsics	Last-digit differences on string output, platform-dependent differences on transcendentals
Raw byte semantics of UTF8String (invalid UTF-8) (#4764)	Arrow requires valid UTF-8, so byte-preserving behavior cannot be reproduced
Lambda bodies or arbitrary child expression trees (transform, aggregate, zip_with)	The native path would need to evaluate an arbitrary Catalyst tree, not a fixed kernel
Dependence on JVM session state beyond a scalar config (SQLConf lookups, user classes, reflection)

Rate it

• High: no hazard applies, or every hazard that applies is fully specified and mechanically reproducible (for example integer arithmetic with a documented overflow rule). A reviewer reading the Spark source could enumerate every case.
• Medium: hazards apply but are bounded and detectable from the expression or the batch at runtime, so the incompatible cases can be routed back through the dispatcher via CodegenDispatchFallback while the common case runs natively. Name the guard and say how it is detected. "We will document the difference" is not a guard.
• Low: matching Spark requires reimplementing a JVM library (ICU collation, DecimalFormat), or the incompatible cases cannot be detected before evaluating, or the only honest native path is Incompatible by default.

A Low rating is a hard disqualifier regardless of upside. The project's direction is the opposite of shipping incompatible-by-default native paths (see #4506 and the discussion in #4654). Verify both issues are still open before citing them.

Step 4: Score native upside

Read Spark's doGenCode (and nullSafeEval) for the expression and answer, concretely: what does the JVM allocate or recompute per row, and can a Rust kernel avoid it?

Positive signals, strongest first:

1. Per-row heap allocation in the generated code. Count the allocation sites: new UTF8String, UTF8String.fromString, .getBytes(), Decimal.apply, new GenericArrayData, new StringBuilder, boxing. Multiply by batch size (8192) to state the allocation the native path removes. This is the most defensible upside argument, and it is the one that maps directly to the "reduction in memory allocation" question.
2. Work that is constant across the batch but repeated per row. A regex compiled per row, a formatter constructed per row, a literal argument re-parsed per row. Native kernels hoist this (see the technique table).
3. A technique from optimizing_expressions.md that plainly applies. Vectorized unary/binary kernels, zero-copy buffer reuse, Cow borrow, an ASCII byte-offset fast path, a lookup table, direct offset-buffer reads. Name the technique and the shape it helps.
4. Nested or complex input/output types. Per-row reads of ListVector / StructVector / MapVector through the JVM wrappers allocate per row per level. Native code operates on the child arrays and offsets directly.
5. Specific inputs of an in-pool expression that fall back to Spark entirely today (found in Step 2: an input or output type outside isSupportedDataType, so the dispatcher cannot compile the kernel and the whole operator falls back). Recovering those inputs dwarfs removing a JNI round trip. This is about inputs of an expression that is otherwise dispatched, not about expressions Comet does not support at all, which are out of pool.
6. Presence in real workloads. TPC-H / TPC-DS / common ETL shapes. A large relative win on an expression nobody runs is not worth a contributor's week.

Anti-signals, each of which pushes the rating down:

• The generated code is straight primitive arithmetic with no allocation. The per-row JVM cost is already a few nanoseconds and the JNI round trip amortizes over 8192 rows.
• The cost is dominated by data-dependent work where Rust has no structural advantage (parsing a large JSON document, a heavy regex match).
• The result is a nested type that must be built element by element in either language.
• The expression is almost always folded away as a constant, or its argument is a literal that Spark's own optimizer already handles.

Rate it

• High: allocation per row that scales with batch size and a named vectorization or zero-copy technique that removes it, on an expression that appears in mainstream workloads. Or recovery of inputs that fall back to Spark for the whole operator today.
• Medium: real allocation or hoistable per-row work, but on an expression that is uncommon in practice, or where the achievable win is a modest constant factor.
• Low: primitive arithmetic with no allocation, or the win is confined to the per-batch JNI round trip, or the expression is rare.

Step 5: Calibrate empirically when the verdict is borderline

Steps 3 and 4 are a static assessment and need no build. That is deliberate: the skill should be cheap to run over a pool of candidates. But a Medium on either axis is exactly where a static read is least trustworthy, so when either rating is Medium, gather at least one piece of empirical evidence before writing the verdict. Building is required from here (make; see CLAUDE.md).

Any one of these is sufficient, cheapest first:

1. Count allocations in the actual generated kernel. The dispatcher's emitted Java is inspectable without running a query, via CometBatchKernelCodegen.generateSource. Model a scratch test on spark/src/test/scala/org/apache/comet/CometCodegenSourceSuite.scala, dump the body for the candidate over a representative Arrow schema, and count the per-row allocation sites in the real code rather than in Spark's source. Delete the scratch test afterwards, or keep it only if it asserts something durable.

2. Proxy benchmark against a comparable native expression. Pick an already-Native expression with the same arity and types over the same data, and benchmark it against the candidate using the existing JVM harness (CometBenchmarkBase.runExpressionBenchmark, see spark/src/test/scala/org/apache/spark/sql/benchmark/CometStringExpressionBenchmark.scala). The gap between the two is an estimate of the ceiling for the native win. Say plainly that it is a proxy: it bounds the win, it does not predict it.

   SPARK_GENERATE_BENCHMARK_FILES=1 make benchmark-org.apache.spark.sql.benchmark.Comet<Family>ExpressionBenchmark
   

3. Allocation rate on a benchmark run. Add -Xlog:gc (or a JFR recording) to the benchmark JVM and compare total allocated bytes for the candidate against the native comparison expression. Report bytes per row.

Whatever you measure, quote the numbers in the issue and the command that produced them. An unquoted measurement is indistinguishable from a guess.

Step 6: Verdict

Combine the two axes. Effort is a tiebreak, not a gate.

Compatibility	Upside	Verdict
High	High	Recommended. File the issue.
High	Medium	Recommended if the empirical calibration from Step 5 supports it, otherwise Deferred with the numbers.
High	Low	Disqualified. The dispatcher is already exact and the native win would be marginal.
Medium	High	Recommended, scoped to the compatible subset with a CodegenDispatchFallback guard for the rest.
Medium	Medium	Deferred. Record the blocker and what would change the answer.
Medium	Low	Disqualified.
Low	any	Disqualified. Compatibility is a gate, not a trade-off.

Every Deferred and Disqualified verdict must name the specific blocker, in a form a future run can re-check: "the Rust regex crate has no lookaround, and Spark's pattern surface allows it", not "compatibility risk". A vague blocker guarantees the expression gets re-assessed from scratch, which is the exact waste this skill exists to prevent.

Also state the effort estimate (roughly: wiring an upstream function, a self-contained kernel, or a kernel plus shims plus a new native code path) so a contributor can judge whether to pick it up.

Step 7: File the issue

File exactly one issue per run.

• Recommended: file the proposal.
• Deferred: file it too, with the blocker as the headline and the "what would change the answer" condition spelled out. A deferral with a concrete unblocking condition is useful to the project; a deferral that lives only in this conversation is not.
• Disqualified: do not file an issue. The audit-log line in Step 8 is the record. In survey mode, go back to Step 1, take the next-ranked candidate, and assess that one, so the run still ends with one filed issue. In single-expression mode, report the disqualification to the user and stop after Step 8.

Check the labels exist before using them (gh label list --repo apache/datafusion-comet). Use enhancement, area:expressions, and performance, plus the family label when one applies (temporal expressions, array expressions, json expressions, map expressions). Do not create new labels.

Title: Implement <function> natively instead of JVM codegen dispatch. For a deferral: Native candidate assessment: <function> deferred on <blocker>.

Body sections, in this order:

1. How it runs today. The serde, the trait it uses, and what that means at runtime. Note any inputs that fall back to Spark entirely rather than to the dispatcher.
2. Native upside. The Step 4 rating with its evidence: allocation sites per row times batch size, the technique that removes them, workload presence, and any Step 5 measurement with the command that produced it.
3. Compatibility assessment. The Step 3 rating, the hazard checklist result (state which hazards apply and which were checked and cleared), cross-version differences that need shims, and whether an upstream datafusion-spark or datafusion-functions implementation exists.
4. Proposed approach. One of three, depending on the sub-pool and what upstream offers: "wire the upstream function" (point at wire-datafusion-function), "implement a native kernel" (point at implement-comet-expression), or, for sub-pool B, "make the existing native path the default by guarding the divergent inputs" (point at audit-comet-expression, and name the config key that gates it today). For a Medium compatibility rating, state the compatible subset, the runtime guard that detects the rest, and that the rest stays on CodegenDispatchFallback.
5. Non-goals. Anything the issue deliberately leaves out, most often the incompatible subset and any collation work.
6. Acceptance criteria. Bit-identical output for the covered subset including nulls and error behavior, a criterion benchmark in native/spark-expr/benches/ covering the shapes required by optimizing_expressions.md, Comet SQL Tests for the expression, and an updated audit-log line.
7. Footer. A line noting the issue was produced by the suggest-native-expression skill, a link to the "Native Coverage for Codegen-Dispatched Expressions" section of the roadmap, and a pointer to the audit-log page where the assessment is recorded.

Use a temp file for the body to keep quoting sane:

gh issue create --repo apache/datafusion-comet \
  --title "Implement <function> natively instead of JVM codegen dispatch" \
  --label "enhancement,area:expressions,performance" \
  --body-file /tmp/native-candidate-<function>.md

Print the URL.

Step 8: Record the assessment in the audit log

This step is not optional, and it runs for every verdict including Disqualified. It is what makes the pool shrink over time instead of being re-walked on every run.

Add the line to the expression's category page under docs/source/contributor-guide/expression-audits/, named after the Spark function-registry category (string_funcs.md, datetime_funcs.md, and so on). Match the category used for the expression in docs/source/user-guide/latest/expressions.md. If the page does not exist, create it with the ASF license header, a # <category> Expression Audits title, and the standard intro blockquote the other pages use. Add or reuse a ## <function_name> heading, keeping headings alphabetically ordered.

Get the date from the system, not from memory:

date -u +%Y-%m-%d

One line, in this shape:

- Native candidate (assessed 2026-08-13): Disqualified. Compatibility: High (pure integer
  arithmetic, no locale or timezone dependence). Upside: Low (generated code allocates nothing
  per row; the only saving is the per-batch JNI round trip). Blocker: no measurable win to
  justify a native kernel. No issue filed.

- Native candidate (assessed 2026-08-13): Recommended ([#5350](https://github.com/apache/datafusion-comet/issues/5350)).
  Compatibility: High (upstream `datafusion-spark` implementation matches Spark for all input
  types; no collation surface). Upside: High (two `UTF8String` allocations per row, 16k per
  batch, removable with a zero-copy `Cow` borrow; appears in TPC-DS q13).

Rules for the line:

• Always include the verdict, both ratings with a parenthesized one-clause reason each, and either the issue link or No issue filed.
• Deferred and Disqualified lines must carry a Blocker: clause naming the specific thing that would have to change.
• Never invent an issue number. Use the URL that gh issue create printed, or verify an existing one with gh issue view <N> --repo apache/datafusion-comet.
• Never edit or delete a previous Native candidate line. A re-assessment adds a new dated line below it, and says what changed since the last one.
• Keep it to one bullet. These pages are scanned, not read.

Then format the markdown, because CI checks it:

npx prettier "docs/source/contributor-guide/expression-audits/*.md" --write

Commit the audit-log change on a branch and open a PR for it (the repo's PR template applies). The filed issue and the audit-log line are two halves of one deliverable: the issue is the call to action, the log line is the durable record that the call was made.

Red flags: stop and re-read

• "Native is obviously faster than JVM." The dispatcher already runs in-pipeline with zero-copy string reads and a compile-once kernel. If you cannot name the per-row allocation or the vectorization technique, the upside rating is Low and the verdict is Disqualified.
• "We can match Spark for the common cases." Then the compatibility rating is Medium at best, and the proposal must name the runtime guard and keep the rest on CodegenDispatchFallback. Replacing an exact path with an approximate one is a regression, not an optimization.
• "The expression has no audit entry, so nobody has looked at it." Run the Step 1 grep. A Disqualified line means somebody did look, and re-proposing it without addressing the recorded blocker wastes a reviewer's time.
• "I will estimate the speedup." An estimate presented as a measurement is the failure mode that produces abandoned issues. Either measure it (Step 5) or label it explicitly as a static estimate in both the issue and the log line.
• "This is a big win, I will file issues for the whole family." One issue per run. A family sweep needs a project-direction discussion, not a batch of generated issues.

Rationalization table

Excuse	Reality
"Recording a disqualification is busywork, nobody reads the log"	The log is the only thing that stops the next run from re-assessing the same expression. An unrecorded disqualification is a repeated cost.
"The blocker is obvious from the discussion"	The discussion evaporates when the session ends. A blocker that is not in the log line does not exist.
"Upstream probably has it, I will assume it matches Spark"	datafusion-spark functions are not all complete. Read the implementation and diff the semantics before crediting it.
"Compatibility can be sorted out during implementation"	Compatibility is the gate. An issue that hides a Low rating sends someone down a path the project will not merge.
"I will file the issue and skip the audit-log PR"	Then the pool never shrinks. Both deliverables or neither.

Related skills

• wire-datafusion-function — the follow-through when upstream already has the function.
• implement-comet-expression — the follow-through when a new native kernel is needed.
• optimize-comet-expression — for an expression that is already native and merely slow.
• audit-comet-expression — the mechanics of closing a compatibility gap on a sub-pool B (Hybrid) expression, once this skill has established the win is worth it.

Tone and style

• Show the ranking, the checklist result, and the numbers. A verdict without its evidence cannot be challenged, and an unchallengeable verdict is not useful.
• Use backticks around function names, config keys, file paths, types, and class names.
• Be explicit about what is measured versus estimated.
• Avoid em dashes and semicolons; use separate sentences instead.