macrame-db 0.15.0

//! What the write actor knows about its own latency (T1.4, D-079).
//!
//! # Why this exists
//!
//! [`crate::CHUNK_BUDGET`] is 3 ms and the crate has, until now, had exactly one
//! way to find out whether that bound holds: run `benches/budgets.rs` on a
//! synthetic fixture. That is a statement about a laptop, not about a database
//! in use. D-059 already established that the bound does **not** hold on a large
//! file, by a factor of 15, and it took a benchmark rewrite to notice — because
//! nothing in the running system was counting.
//!
//! Tier 1's other three items are all "make the tail bounded". None of them can
//! be validated in the field without something that measures the tail, which is
//! why this is a precondition for them rather than a nice-to-have.
//!
//! # What is recorded, and what is deliberately not
//!
//! Four things, all of them per **actor turn** — one command, start to finish:
//!
//! - **queue depth** on both channels, sampled *before* the turn begins;
//! - **hold duration**, bucketed, per command kind;
//! - **holds over budget**, counted separately per kind;
//! - **the longest hold since open**, with the kind that caused it.
//!
//! The hold is the whole turn, not the `execute` call's SQL. That is the
//! quantity the budget is about: the SQLite write lock is not preemptible, so an
//! interactive assertion arriving mid-turn waits for the turn, whatever the turn
//! spent its time on.
//!
//! There is no per-command timestamp trail and no sampling of individual slow
//! commands. That would be a tracing problem, and `tracing` is already a
//! dependency — spans belong there. This module answers one question ("is the
//! bound holding, and if not, which kind breaks it") in fixed memory, with no
//! allocation on the actor's path.
//!
//! # The feature gate
//!
//! Behind `metrics`, which has been a **default** feature since 0.12.11
//! (D-154): a crate whose contract is a latency bound must not ship a default
//! build that cannot report whether the bound is met. `--no-default-features`
//! still removes it. With the feature off, [`ActorMetrics`] is a
//! zero-sized type whose methods compile away and [`HoldTimer::start`] does not
//! read the clock — so the actor loop has **one** shape either way. That
//! matters more than the nanoseconds: a `#[cfg]` in the loop body is how the
//! instrumented and uninstrumented paths drift until only one of them is the one
//! that runs.

use std::time::Duration;

/// The command kinds the actor can spend a turn on.
///
/// One flat enum across both channels rather than one per channel. The question
/// this exists to answer is "which command broke the budget", and a reader
/// looking at a 400 ms hold does not first want to know which queue it came off.
/// Priority is a property of scheduling; kind is a property of cost.
///
/// # `#[non_exhaustive]`, added while it was still free (0.12.8, W4.2)
///
/// Adding a variant here is a **breaking change** without this attribute,
/// because a downstream `match` on `CommandKind` would stop compiling. That is
/// not hypothetical for this enum: [`crate::metrics::CommandKind::Rehydrate`]
/// did not exist until 0.12.9 precisely because adding it was a break, and
/// rehydration reported as `Archive` for several releases as a result. The
/// codebase has already paid this cost once, which is the argument for paying
/// the attribute now rather than deciding it at 1.0 when the cost is permanent.
///
/// Callers must therefore include a `_ =>` arm. In exchange, this enum can grow
/// a variant for a command kind that does not exist yet without a major version.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
#[repr(u8)]
#[non_exhaustive]
pub enum CommandKind {
    AssertEdge,
    RetireEdge,
    UpsertConcept,
    WriteBulkAtomic,
    RebuildCurrent,
    RegisterModel,
    Shutdown,
    BulkImportChunk,
    WriteConceptsChunk,
    WriteAnalyticsChunk,
    UpsertEmbeddingChunk,
    Archive,
    RebuildFts,
    /// The **fill** half of a chunked shadow rebuild — `Begin` and every
    /// `Fill` chunk (T1.2).
    ///
    /// Its own kind rather than folded into `RebuildCurrent`, because the two
    /// have opposite latency profiles and the whole point of the chunked path
    /// is that its turns are short — averaging them together would hide
    /// exactly the improvement.
    ///
    /// **Fill-only since 0.14.16** ([D-233]). Through 0.14.15 this kind also
    /// carried the swap turn, which is over budget by construction, so its
    /// `over_budget` count was `N(rebuilds) + regressions` and could not be
    /// decomposed — the counter was a constant, not a signal. The swap is
    /// [`CommandKind::ShadowSwap`] now, and what is left here is the half that
    /// is *meant* to fit [`crate::CHUNK_BUDGET`]. A nonzero count on this kind
    /// is therefore a clean canary: a fill chunk ran long, which is a
    /// regression and nothing else.
    ///
    /// [D-233]: ../docs/architecture/s13-decision-register.md#d-233
    ShadowRebuild,
    /// Refreshing the query planner's statistics (0.12.4, D-149).
    ///
    /// Its own kind rather than folded into `RebuildFts`, though both are
    /// maintenance on derived state: this one is bounded by
    /// `PRAGMA analysis_limit` and that one is bounded by the size of the
    /// concept table, so averaging their holds together would describe neither.
    Analyze,
    /// Moving archived rows back into the hot file (0.12.9, W4.3, D-152).
    ///
    /// Its own kind at last. Through 0.12.8 this reported as
    /// [`CommandKind::Archive`], on the stated ground that rehydration is the
    /// archive path run backwards and shares its budget — true of the *budget*
    /// and false of the *attribution*, which is what a metrics surface is for.
    /// An operator reading a long `archive` hold could not tell whether the
    /// database had archived anything at all, and the two move rows in opposite
    /// directions.
    ///
    /// The real reason it stayed folded was that adding a variant was a
    /// breaking change. `#[non_exhaustive]` (W4.2) is what removed that
    /// obstacle, and this variant is the first thing it bought — which is also
    /// the evidence that the attribute was worth adding rather than a
    /// precaution against a hypothetical.
    ///
    /// **Appended at the end**, per [`CommandKind::index`]: the position of
    /// every existing variant is a persisted contract in two languages.
    Rehydrate,
    /// An explicit `PRAGMA wal_checkpoint` (0.12.13, W5.2, D-156).
    ///
    /// Its own kind because it is the one actor turn that is **not** a
    /// transaction: it moves frames from the WAL back into the main database
    /// file, and its duration is a function of how much WAL has accumulated
    /// rather than of anything the caller passed. Folding it into any existing
    /// kind would make that kind's hold distribution bimodal for a reason no
    /// dashboard could recover.
    ///
    /// **Appended at the end**, per [`CommandKind::index`].
    Checkpoint,
    /// `PRAGMA optimize` — re-analysing only what SQLite believes has drifted
    /// (0.13.24, W10.5, D-197).
    ///
    /// Split out of [`CommandKind::Analyze`], which covered both from 0.12.4 to
    /// 0.13.23. The split is [`CommandKind::Rehydrate`]'s lesson applied before
    /// the fact rather than after it: [D-168] refused to decide `Analyze`'s
    /// budget exemption *because* the kind was shared, since a judgement made
    /// about the explicit call would have landed on the automatic one —
    /// `close()` runs `optimize()` unconditionally — without ever being made
    /// about it.
    ///
    /// The two also have genuinely different hold distributions, which is the
    /// same argument [`CommandKind::ShadowRebuild`] is separate on.
    /// [`crate::Database::analyze`] does the work unconditionally and its hold
    /// tracks the table. This one is a no-op when nothing has moved, so its
    /// distribution is bimodal by design and averaging the two together
    /// describes neither.
    ///
    /// **Appended at the end**, per [`CommandKind::index`].
    ///
    /// [D-168]: ../docs/architecture/s13-decision-register.md#d-168
    Optimize,
    /// Registering a lineage (0.14.7, §15.4).
    ///
    /// Its own kind rather than folded into `AssertEdge`, though both are one
    /// small transaction: a fork writes to `branches` and nothing else, so its
    /// hold is the floor an actor turn can have, and averaging it into a
    /// command that touches four tables would flatter that command's numbers.
    ///
    /// Last in declaration order because that order is a persisted contract and
    /// **new variants go at the end** — see [`CommandKind::index`]. Grouping it
    /// next to `RegisterModel`, which is where it belongs by kind, would have
    /// renumbered nine counters and relabelled the Python histogram's axes.
    Fork,
    /// Forgetting a lineage (0.14.13, §15.4, D-230).
    ///
    /// Its own kind rather than folded into [`CommandKind::Archive`], on
    /// [D-152]'s finding rather than on a fresh argument: the budget really is
    /// shared and the attribution is not, and an operator reading a long
    /// `archive` hold could not tell whether the database had archived a
    /// backlog of closed intervals or dropped an abandoned branch. The two also
    /// have unrelated cost curves — one is a function of how long it has been
    /// since the last run, the other of how much was written on one branch.
    ///
    /// At the end of the declaration order, per [`CommandKind::index`].
    ///
    /// [D-152]: ../docs/architecture/s13-decision-register.md#d-152
    ArchiveBranch,
    /// The **swap** turn of a chunked shadow rebuild (0.14.16, D-233).
    ///
    /// Split out of [`CommandKind::ShadowRebuild`], which covered both halves
    /// from 0.6.0 to 0.14.15. This is the third instance of one shape —
    /// [`CommandKind::Rehydrate`] out of `Archive` ([D-152]),
    /// [`CommandKind::Optimize`] out of `Analyze` ([D-197]), this — so the
    /// class is named where it can be seen: **one `CommandKind`, one
    /// structural hold distribution.** A kind covering two is a defect on
    /// arrival, to be split in review rather than found by probe.
    ///
    /// Here the bimodality is structural rather than workload-dependent, which
    /// is what makes it the clearest instance of the three. Index names are
    /// global and SQLite has no `ALTER INDEX … RENAME`, so the shadow cannot
    /// carry `idx_lc_traversal_cover` while the live table still holds that
    /// name — the swap is where all three indexes get built, under the write
    /// lock. [D-082](../docs/architecture/s13-decision-register.md#d-082)
    /// measured it at **46.8 ms**, 15.6× the budget, and it grows with the
    /// table.
    ///
    /// **Exempt**, unlike its fill half — see
    /// [`CommandKind::exempt_from_budget`], where the criterion is stated.
    ///
    /// At the end of the declaration order, per [`CommandKind::index`].
    ///
    /// [D-197]: ../docs/architecture/s13-decision-register.md#d-197
    ShadowSwap,
}

impl CommandKind {
    /// Every kind, in declaration order. Indexing into the per-kind arrays is by
    /// position in this slice, so the two must not drift — which is why the
    /// arrays are sized from `ALL.len()` rather than from a hand-written count.
    pub const ALL: &'static [CommandKind] = &[
        CommandKind::AssertEdge,
        CommandKind::RetireEdge,
        CommandKind::UpsertConcept,
        CommandKind::WriteBulkAtomic,
        CommandKind::RebuildCurrent,
        CommandKind::RegisterModel,
        CommandKind::Shutdown,
        CommandKind::BulkImportChunk,
        CommandKind::WriteConceptsChunk,
        CommandKind::WriteAnalyticsChunk,
        CommandKind::UpsertEmbeddingChunk,
        CommandKind::Archive,
        CommandKind::RebuildFts,
        CommandKind::ShadowRebuild,
        CommandKind::Analyze,
        CommandKind::Rehydrate,
        CommandKind::Checkpoint,
        CommandKind::Optimize,
        CommandKind::Fork,
        CommandKind::ArchiveBranch,
        CommandKind::ShadowSwap,
    ];

    pub const COUNT: usize = CommandKind::ALL.len();

    /// This kind's slot in the per-kind arrays.
    ///
    /// # Declaration order is a persisted contract (0.12.8, W4.2)
    ///
    /// `self as usize` means the **order of the variants above** is the order of
    /// every per-kind array in this module, and the compiler cannot catch a
    /// change to it. Reordering the enum silently reassigns every counter to a
    /// different command: the code compiles, the tests pass, and a histogram
    /// read after the change attributes `archive`'s holds to `rebuild_fts`.
    ///
    /// **New variants go at the end**, always — including at the end of
    /// [`CommandKind::ALL`], whose order is what `as_str()` and the Python
    /// surface enumerate. This binds Python too: `BUCKET_BOUNDS_MICROS` is a
    /// module constant there and `KindMetrics` is built by position, so a
    /// reorder here relabels axes in a language the Rust compiler is not
    /// looking at.
    ///
    /// `#[repr(u8)]` is on the enum for the same reason — it pins the
    /// discriminants to the declaration order rather than leaving them to the
    /// compiler — but it pins them to whatever the order *is*, so it does not
    /// make a reorder safe. Only this rule does.
    pub const fn index(self) -> usize {
        self as usize
    }

    pub const fn as_str(self) -> &'static str {
        match self {
            CommandKind::AssertEdge => "assert_edge",
            CommandKind::RetireEdge => "retire_edge",
            CommandKind::UpsertConcept => "upsert_concept",
            CommandKind::WriteBulkAtomic => "write_bulk_atomic",
            CommandKind::RebuildCurrent => "rebuild_current",
            CommandKind::RegisterModel => "register_model",
            CommandKind::Shutdown => "shutdown",
            CommandKind::BulkImportChunk => "bulk_import_chunk",
            CommandKind::WriteConceptsChunk => "write_concepts_chunk",
            CommandKind::WriteAnalyticsChunk => "write_analytics_chunk",
            CommandKind::UpsertEmbeddingChunk => "upsert_embedding_chunk",
            CommandKind::Archive => "archive",
            CommandKind::RebuildFts => "rebuild_fts",
            CommandKind::ShadowRebuild => "shadow_rebuild",
            CommandKind::Analyze => "analyze",
            CommandKind::Rehydrate => "rehydrate",
            CommandKind::Checkpoint => "checkpoint",
            CommandKind::Optimize => "optimize",
            CommandKind::Fork => "fork",
            CommandKind::ArchiveBranch => "archive_branch",
            CommandKind::ShadowSwap => "shadow_swap",
        }
    }

    /// Whether this kind is exempt from [`crate::CHUNK_BUDGET`] by contract.
    ///
    /// The exemptions are the table in `CHUNK_BUDGET`'s own rustdoc, and they
    /// are carried here so a dashboard can separate "the budget is being
    /// broken" from "the budget does not apply and never claimed to". Counting
    /// an `archive` as a budget violation would make the violation count useless
    /// on any database that archives.
    ///
    /// The two lists must agree, and since 0.12.9 they are tied together in
    /// both directions by `the_budget_exemptions_and_their_documented_table_agree`
    /// — the extra-row direction being the one worth having, since a table row
    /// with no code behind it promises a caller an exemption the violation
    /// counter is about to disagree with.
    ///
    /// # The criterion, stated at last (0.14.16, W12.16, [D-233])
    ///
    /// The register applied one rule three times without naming it, and naming
    /// it is what let the fourth case be decided rather than argued.
    ///
    /// > **Exempt means the chunk bound does not apply: the operation is atomic
    /// > by necessity and has no smaller unit. Counted means the bound applies,
    /// > so exceeding it is information.**
    ///
    /// Every exemption on this list was argued that way in its own release,
    /// whatever the summary sentence said afterwards.
    /// [`CommandKind::WriteBulkAtomic`] ([D-014]) is one statement, and is the
    /// kind that exists precisely because the chunked variant is *not* atomic.
    /// [`CommandKind::Archive`] ([D-012]) and [`CommandKind::ArchiveBranch`]
    /// delete a consistent set or none of it.
    /// [`CommandKind::RebuildCurrent`] ([D-023]) re-derives a whole projection
    /// in one transaction. [`CommandKind::Rehydrate`] ([D-152]) is one
    /// unchunked transaction moving rows back across the file boundary.
    /// [`CommandKind::Checkpoint`] ([D-156]) is a WAL boundary. None of them
    /// has a smaller unit to chunk *into*, so 3 ms is not a bound they failed —
    /// it is a bound that was never about them.
    /// [`CommandKind::Analyze`] and [`CommandKind::Optimize`] ([D-197]) do have
    /// one: they are bounded work that can take longer or shorter, so exceeding
    /// is a fact about this database and worth counting.
    ///
    /// # The criterion took three tries, and the two that failed are the useful part
    ///
    /// **v1 — *expected-on-healthy is exempt, workload-dependent is not*.**
    /// Falsified by reading [D-197] closely rather than by any new measurement:
    /// `Optimize` **runs on every close** and stays counted. If expectedness
    /// decided the question, `Optimize` would be exempt and it is not.
    ///
    /// **v2 — *if `over_budget` can differ from `turns`, count it*.** Falsified
    /// by three of the exemptions themselves: an `Archive` with nothing
    /// archivable, a `Rehydrate` of a single row and a `Checkpoint` on an empty
    /// WAL all come in *under* budget, so their counters can differ from their
    /// turn counts and the rule would un-exempt all three.
    ///
    /// Both were **observational** — read off the outcomes the existing
    /// exemptions happened to produce, and so decidable only after the fact.
    /// Inapplicability is decidable at design time from what the operation *is*,
    /// which is what a criterion has to be if it is to settle the next case
    /// rather than rationalise the last one.
    ///
    /// Expected-on-healthy survives as **corroboration, not definition**: a kind
    /// with no smaller unit usually does exceed on every healthy database, so
    /// the symptom is a fair sanity check on the diagnosis. `Optimize` is
    /// exactly the case that shows why it cannot be the test itself.
    ///
    /// `over_budget` is incremented once per turn that exceeds, so it counts
    /// **occurrences and not magnitude**. That is the fact the criterion turns
    /// on: a kind whose every turn exceeds contributes a constant to
    /// [`MetricsSnapshot::budget_violations`] and moves not at all when the
    /// hold doubles. Growth is visible in this kind's histogram and
    /// [`KindSnapshot::longest`], which no exemption touches.
    ///
    /// # The two halves of a shadow rebuild land on opposite sides of it
    ///
    /// [`CommandKind::ShadowSwap`] is exempt: ≥ 15.6× by construction ([D-082]
    /// measured 46.8 ms against a 3 ms budget), with no healthy state in which
    /// it fits, and routine — the crate's own end-to-end suite triggers one.
    /// Counting it would put a permanent `N(rebuilds)` in the violation list
    /// of every database that has ever repaired its projection, which is
    /// [`CommandKind::Rehydrate`]'s argument exactly.
    ///
    /// [`CommandKind::ShadowRebuild`] — the fill half — is **not** exempt, and
    /// that is the half [D-082] was protecting when it refused to exempt the
    /// merged kind: *"exempting the kind would hide the first fact to excuse
    /// the second."* The goal is reaffirmed and the mechanism superseded. The
    /// split protects fill structurally, where non-exemption of the merged
    /// kind only protected it in principle: a fill regression used to arrive
    /// as `+1` on a counter that already read `N(rebuilds)`, and now it is the
    /// only thing that can move `shadow_rebuild` off zero at all.
    ///
    /// `a_swap_over_budget_is_not_a_violation` is what keeps this honest, and
    /// its fixture is the load-bearing part: it seeds enough of a graph to put
    /// the swap **over** the budget, asserts that first, and only then asserts
    /// the swap's own count is zero. The obvious form — run a rebuild, assert
    /// the violation list is empty — is worthless twice over. On a small
    /// fixture the swap finishes inside 3 ms and the assertion passes whether
    /// the kind is exempt or not; on a real one the *fill* chunks exceed the
    /// budget legitimately (3.14 ms at 200 keys in a debug build), so an empty
    /// list is a property of small fixtures rather than of rebuilds.
    ///
    /// The two tests are **one instrument with two asymmetric halves**, and it
    /// is worth being exact about which owns what.
    /// `a_swap_over_budget_is_not_a_violation` owns **narrowing**: re-count the
    /// swap and its assertion moves off zero. It cannot own widening, because
    /// widening an exemption only ever *removes* entries from
    /// [`MetricsSnapshot::budget_violations`] — an assertion that a count is
    /// zero stays green under every widening, including one that swallows the
    /// fill half whole. **Widening is owned by
    /// `a_long_fill_is_a_violation_and_a_long_swap_is_not` below and by nothing
    /// else**, because forging a long fill and asserting it **is** counted is
    /// the only shape of assertion a widening can break.
    ///
    /// # Any claim about fill and this budget must name a build mode and a fixture size
    ///
    /// At fixture scale the 3 ms bound sits **inside** fill variance rather
    /// than above it, so the same assertion is true or false depending on how
    /// the binary was compiled and how much graph it was handed. Debug
    /// especially: 200 keys × 4 generations puts the longest fill at 3.14 ms —
    /// one violation, the counter working — while a release build of the same
    /// shape stays under. A test that asserts anything about fill overages is
    /// therefore asserting something about *its own fixture and profile*, and
    /// has to say which. The swap is the opposite and that is why the exemption
    /// is testable at all: it exceeds by 15.6× and it exceeds in every mode.
    ///
    /// [D-012]: ../docs/architecture/s13-decision-register.md#d-012
    /// [D-014]: ../docs/architecture/s13-decision-register.md#d-014
    /// [D-023]: ../docs/architecture/s13-decision-register.md#d-023
    /// [D-082]: ../docs/architecture/s13-decision-register.md#d-082
    /// [D-152]: ../docs/architecture/s13-decision-register.md#d-152
    /// [D-156]: ../docs/architecture/s13-decision-register.md#d-156
    /// [D-197]: ../docs/architecture/s13-decision-register.md#d-197
    /// [D-233]: ../docs/architecture/s13-decision-register.md#d-233
    ///
    /// # `Rehydrate` is exempt, and splitting it out is what made that a
    /// decision rather than an accident (0.12.9, W4.3, D-152)
    ///
    /// Until 0.12.8 rehydration reported as [`CommandKind::Archive`] and was
    /// therefore exempt **by inheritance** — nobody had decided it, it fell out
    /// of the borrowed kind. Giving it its own variant would have silently
    /// flipped it to non-exempt, and since a rehydrate is one unchunked
    /// transaction moving rows back across the file boundary, every single one
    /// would have counted as a budget violation. The violation count would then
    /// have become useless on any database that rehydrates, which is precisely
    /// the failure the `Archive` exemption exists to prevent, arriving by the
    /// back door of a change made for attribution.
    ///
    /// So it is exempt, on the merits and now on the record: rehydration is the
    /// archive path run backwards and makes the same claim about its hold —
    /// that it is bulk movement with no latency bound, and that the caller asked
    /// for it explicitly.
    ///
    /// # Neither [`CommandKind::Analyze`] nor [`CommandKind::Optimize`] is
    /// exempt, and since 0.13.24 those are two decisions (W10.5, D-197)
    ///
    /// They were one kind from 0.12.4 to 0.13.23, and [D-168] declined to decide
    /// the exemption *because* they were: `Analyze` covered
    /// [`crate::Database::optimize`] too, `close()` calls that unconditionally,
    /// and so a judgement made about the explicit call would have landed on the
    /// automatic one without ever being made about it. That is
    /// [`CommandKind::Rehydrate`]'s lesson above arriving from the other
    /// direction — there a shared kind *granted* an exemption nobody had
    /// decided; here one would have *laundered* one. W10.5 split the kind so
    /// each could be answered on its own evidence. Both answers came back the
    /// same and the reasons are different, which is the whole reason the split
    /// had to come first.
    ///
    /// **[`CommandKind::Analyze`] cannot state a `Bound`, so it cannot have a
    /// row.** `ANALYZE` is one indivisible statement whose cost is set by data
    /// volume — measured at **5.26 ms at 10,000 edges and 19.1 ms at 40,000**
    /// against a 3 ms budget (`examples/analyze_hold.rs`, [D-166]). Every call
    /// is a violation and always will be. `Checkpoint`'s bound is frames
    /// accumulated since the last one; `Archive`'s is the session's row count.
    /// The honest entry here would be "the size of the table, damped 3–4× by
    /// `analysis_limit`", which is not a bound but the absence of one, and a
    /// row that cannot fill that column is this table admitting the thing it
    /// exists to prevent.
    ///
    /// **[`CommandKind::Optimize`] is not exempt for the opposite reason: its
    /// violations are rare and they are the informative ones.** Measured
    /// (`examples/optimize_hold.rs`, 40,000 edges): **10.7 ms the first time on
    /// a database that has never been analysed, and 90–220 µs every time
    /// after** — comfortably inside the budget, including immediately after a
    /// bulk load that doubled the ledger. It is over budget only when it
    /// actually re-analyses something, and then it is over by a lot: **460 ms**
    /// once the table had grown 25× and SQLite's staleness ratio finally
    /// fired. So the count is bimodal by construction and it is *reporting*
    /// rather than complaining: an `optimize` in `budget_violations()` marks
    /// the calls that did work, which is exactly what an operator wants to
    /// know and exactly what exempting the kind would delete.
    ///
    /// **The violations are expected and must not be "fixed" by lowering
    /// [`crate::schema::ddl::ANALYSIS_LIMIT`].** That would buy the number by
    /// sampling too little to separate the two `source_id`-leading indices,
    /// which is the entire purpose of having statistics ([D-149]).
    ///
    /// [D-149]: ../docs/architecture/s13-decision-register.md#d-149
    /// [D-166]: ../docs/architecture/s13-decision-register.md#d-166
    /// [D-168]: ../docs/architecture/s13-decision-register.md#d-168
    pub const fn exempt_from_budget(self) -> bool {
        matches!(
            self,
            CommandKind::WriteBulkAtomic
                | CommandKind::Archive
                | CommandKind::RebuildCurrent
                | CommandKind::Rehydrate
                | CommandKind::ArchiveBranch
                | CommandKind::Checkpoint
                | CommandKind::ShadowSwap
        )
    }
}

impl std::fmt::Display for CommandKind {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.write_str(self.as_str())
    }
}

/// Upper bounds of the hold-duration histogram, in microseconds.
///
/// `3_000` is [`crate::CHUNK_BUDGET`] exactly, so the bucket boundary and the
/// bound are the same number and a reader does not have to interpolate to answer
/// "what fraction of turns fit". The tail runs to 1 s because D-059's measured
/// worst case was 45 ms and `rebuild_current` at 40K rows is 318 ms (D-077) —
/// a range this has to cover without saturating.
///
/// Anything above the last bound lands in the overflow bucket, which is why
/// [`KindSnapshot::buckets`] is one longer than this slice.
pub const BUCKET_BOUNDS_MICROS: &[u64] = &[
    100, 300, 1_000, 3_000, 10_000, 30_000, 100_000, 300_000, 1_000_000,
];

/// Number of histogram buckets, including the overflow bucket.
pub const BUCKET_COUNT: usize = BUCKET_BOUNDS_MICROS.len() + 1;

#[allow(dead_code)] // used by `imp` under `metrics`, and by the tests always
fn bucket_of(micros: u64) -> usize {
    // Linear scan over nine bounds. A binary search here would be slower in
    // practice and this runs once per actor turn, against a turn measured in
    // microseconds at best.
    BUCKET_BOUNDS_MICROS
        .iter()
        .position(|&bound| micros <= bound)
        .unwrap_or(BUCKET_BOUNDS_MICROS.len())
}

/// Times one actor turn.
///
/// # This clock is no longer optional (0.12.0, W1)
///
/// Until 0.11.0 the field was `#[cfg(feature = "metrics")]` and `elapsed()`
/// returned `Duration::ZERO` in a default build: the reading existed only to
/// feed [`ActorMetrics::record_hold`]'s histogram, so a build that did not keep
/// the histogram had no reason to read a clock.
///
/// The chunk loop changed what the reading is *for*. A chunk's measured hold is
/// now the input to the next chunk's size (`connection::next_chunk_size`, named
/// in prose because it is private — D-144), which means it is a control signal
/// in every build and not an observation in some of them. Left gated, `bulk_import` would have sized its chunks off
/// `Duration::ZERO` — a value that reads as "comfortably under budget" — and
/// grown every chunk to the ceiling, in exactly the builds nobody was measuring.
///
/// So the clock is unconditional and **only the histogram is still gated**:
/// `record_hold` remains a no-op without the feature. What that costs is one
/// `Instant::now()` pair per actor turn — tens of nanoseconds against a turn
/// measured in microseconds at best, and the same reasoning §5.1.5 uses to
/// decide that a channel hop is free beside a chunk.
///
/// It stays a type rather than a bare `Instant::now()` in the loop because the
/// ordering guarantee in [`crate::connection`]'s `Turn` is attached to it.
pub struct HoldTimer {
    start: std::time::Instant,
}

impl HoldTimer {
    #[inline]
    pub fn start() -> Self {
        Self {
            start: std::time::Instant::now(),
        }
    }

    #[inline]
    pub fn elapsed(&self) -> Duration {
        self.start.elapsed()
    }
}

// ---------------------------------------------------------------------------
// Instrumented implementation
// ---------------------------------------------------------------------------

#[cfg(feature = "metrics")]
mod imp {
    use super::{bucket_of, CommandKind, BUCKET_COUNT};
    use std::sync::atomic::{AtomicU64, Ordering};
    use std::time::Duration;

    /// One kind's counters. All `Relaxed`: these are statistics, and ordering
    /// them against each other would buy a consistency no reader needs and cost
    /// fences on the write path.
    #[derive(Debug, Default)]
    struct Kind {
        turns: AtomicU64,
        total_micros: AtomicU64,
        over_budget: AtomicU64,
        /// This kind's own high-water mark, in µs.
        ///
        /// Not redundant with the global `longest`. That one names a single
        /// command, so on any real database it names whichever kind is slowest
        /// overall — and the question "did windowing shrink the archive's worst
        /// hold" cannot be answered by a counter that a bulk import wins. No
        /// packing needed here: the kind is the array index.
        longest_micros: AtomicU64,
        buckets: [AtomicU64; BUCKET_COUNT],
    }

    /// Live counters, shared between the actor and the handle.
    ///
    /// Fixed size, no allocation, no lock. The actor updates; anyone may read.
    #[derive(Debug, Default)]
    pub struct ActorMetrics {
        kinds: [Kind; CommandKind::COUNT],
        /// Packed `micros << 8 | kind`, so the longest hold and the kind that
        /// caused it are read and written as **one** value. Two atomics would
        /// let a reader see a duration from one turn beside a kind from
        /// another — a rare wrong answer to exactly the question this field
        /// exists to answer. 2^56 µs is over two thousand years.
        ///
        /// **The duration must occupy the high bits.** The update is a
        /// `fetch_max` on the packed word, so whichever field is packed high is
        /// the one being compared. The first version of this had the kind up
        /// there, which made the "longest hold" the hold with the largest
        /// *enum index* — a 3 ms `write_concepts_chunk` beat a 10 ms
        /// `rebuild_current` because its variant is declared later. It was
        /// `actor_metrics_tests` that caught it, not the unit tests, because
        /// nothing in the arithmetic is wrong: the packing is only incorrect in
        /// the presence of the atomic operation it exists to serve.
        longest: AtomicU64,
        /// Loop iterations, which is **not** the number of turns taken.
        ///
        /// The depth sample happens at the top of the loop, before `select!`
        /// blocks — so an idle actor has already counted the iteration for a
        /// command that has not arrived. That is right for depth (the sample is
        /// "what was queued when I went looking") and wrong for turns, which is
        /// why [`MetricsSnapshot::turns`] is the sum of the per-kind counters
        /// instead. Conflating the two made `turns` permanently one too high and
        /// disagree with its own breakdown.
        depth_samples: AtomicU64,
        high_depth_sum: AtomicU64,
        high_depth_max: AtomicU64,
        low_depth_sum: AtomicU64,
        low_depth_max: AtomicU64,
        /// Turns where the actor took high-priority work while low-priority
        /// work was already queued (0.12.10, W4.4, D-153).
        ///
        /// The `biased` `select!` in `run_writer_actor` has **no floor**:
        /// sustained high-priority traffic can hold the low tier off
        /// indefinitely, and nothing has ever said whether that happens. This
        /// is the numerator of that question — how often the choice went
        /// against the low tier at all.
        low_starved_turns: AtomicU64,
        /// The current unbroken run of such turns. Reset to zero the moment
        /// low-priority work is taken.
        ///
        /// Not exposed; it is the state [`Self::low_starved_run_max`] is a
        /// high-water mark of. A live value would be read at an arbitrary point
        /// in a run and mean nothing.
        low_starved_run: AtomicU64,
        /// The longest such run since open, which is the number that answers the
        /// question.
        ///
        /// A large `low_starved_turns` on a busy database is unremarkable — it
        /// says the high tier is being used, which is what the tier is for. A
        /// large *run* says one specific low-priority command waited that many
        /// turns, and it is the only one of the two that can distinguish
        /// "prioritised" from "starved".
        low_starved_run_max: AtomicU64,
    }

    const MICROS_SHIFT: u32 = 8;
    const KIND_MASK: u64 = (1 << MICROS_SHIFT) - 1;

    impl ActorMetrics {
        pub fn new() -> Self {
            Self::default()
        }

        /// Sample both queue depths. Called before the turn, not after: after
        /// the turn the queue reflects what arrived *during* it, which is a
        /// different and much less useful quantity.
        #[inline]
        pub fn record_turn(&self, high_depth: usize, low_depth: usize) {
            self.depth_samples.fetch_add(1, Ordering::Relaxed);
            for (sum, max, depth) in [
                (
                    &self.high_depth_sum,
                    &self.high_depth_max,
                    high_depth as u64,
                ),
                (&self.low_depth_sum, &self.low_depth_max, low_depth as u64),
            ] {
                sum.fetch_add(depth, Ordering::Relaxed);
                max.fetch_max(depth, Ordering::Relaxed);
            }
        }

        /// Record which tier the `select!` chose, and what was waiting.
        ///
        /// `low_queued` is the depth sampled *before* the `select!`, so it is
        /// the backlog the turn found on arrival. By the time a high-priority
        /// arm fires the low queue may have grown; using the pre-select reading
        /// keeps this consistent with every other depth figure in this module
        /// and makes the counter conservative — it never invents starvation
        /// from work that arrived after the choice was made.
        ///
        /// A low-priority turn resets the run rather than decrementing it: the
        /// question is "how many turns did one low-priority command wait", and
        /// that is a run length, not a balance.
        #[inline]
        pub fn record_priority_choice(&self, took_high: bool, low_queued: usize) {
            if took_high && low_queued > 0 {
                self.low_starved_turns.fetch_add(1, Ordering::Relaxed);
                let run = self.low_starved_run.fetch_add(1, Ordering::Relaxed) + 1;
                self.low_starved_run_max.fetch_max(run, Ordering::Relaxed);
            } else if !took_high {
                self.low_starved_run.store(0, Ordering::Relaxed);
            }
        }

        #[inline]
        pub fn record_hold(&self, kind: CommandKind, held: Duration) {
            let micros = held.as_micros().min(super::MICROS_CEILING as u128) as u64;
            let k = &self.kinds[kind.index()];
            k.turns.fetch_add(1, Ordering::Relaxed);
            k.total_micros.fetch_add(micros, Ordering::Relaxed);
            k.buckets[bucket_of(micros)].fetch_add(1, Ordering::Relaxed);
            k.longest_micros.fetch_max(micros, Ordering::Relaxed);
            if !kind.exempt_from_budget() && held > crate::CHUNK_BUDGET {
                k.over_budget.fetch_add(1, Ordering::Relaxed);
            }
            self.longest.fetch_max(
                (micros << MICROS_SHIFT) | kind.index() as u64,
                Ordering::Relaxed,
            );
        }

        /// A consistent-enough picture for a dashboard.
        ///
        /// Not a torn-read-free snapshot, and it does not pretend to be: the
        /// actor keeps running while this walks the array, so two kinds may be
        /// read one turn apart. Locking the actor to produce a report would make
        /// the observer a source of the latency it is measuring.
        pub fn snapshot(&self) -> super::MetricsSnapshot {
            let samples = self.depth_samples.load(Ordering::Relaxed);
            let mean = |sum: &AtomicU64| {
                if samples == 0 {
                    0.0
                } else {
                    sum.load(Ordering::Relaxed) as f64 / samples as f64
                }
            };

            let packed = self.longest.load(Ordering::Relaxed);
            let longest_micros = packed >> MICROS_SHIFT;
            let longest = (longest_micros > 0)
                .then(|| {
                    let idx = (packed & KIND_MASK) as usize;
                    CommandKind::ALL
                        .get(idx)
                        .map(|&kind| (kind, Duration::from_micros(longest_micros)))
                })
                .flatten();

            let kinds: Vec<_> = CommandKind::ALL
                .iter()
                .map(|&kind| {
                    let k = &self.kinds[kind.index()];
                    let turns = k.turns.load(Ordering::Relaxed);
                    let total = k.total_micros.load(Ordering::Relaxed);
                    super::KindSnapshot {
                        kind,
                        turns,
                        over_budget: k.over_budget.load(Ordering::Relaxed),
                        mean: total
                            .checked_div(turns)
                            .map_or(Duration::ZERO, Duration::from_micros),
                        longest: Duration::from_micros(k.longest_micros.load(Ordering::Relaxed)),
                        buckets: std::array::from_fn(|i| k.buckets[i].load(Ordering::Relaxed)),
                    }
                })
                .collect();

            super::MetricsSnapshot {
                // Summed, not counted separately — see `depth_samples`.
                turns: kinds.iter().map(|k| k.turns).sum(),
                depth_samples: samples,
                high_depth_mean: mean(&self.high_depth_sum),
                high_depth_max: self.high_depth_max.load(Ordering::Relaxed),
                low_depth_mean: mean(&self.low_depth_sum),
                low_depth_max: self.low_depth_max.load(Ordering::Relaxed),
                low_starved_turns: self.low_starved_turns.load(Ordering::Relaxed),
                low_starved_run_max: self.low_starved_run_max.load(Ordering::Relaxed),
                longest,
                kinds,
            }
        }
    }
}

// ---------------------------------------------------------------------------
// No-op implementation
// ---------------------------------------------------------------------------

#[cfg(not(feature = "metrics"))]
mod imp {
    use super::CommandKind;
    use std::time::Duration;

    /// The `metrics`-off shape: zero-sized, and every method is nothing.
    #[derive(Debug, Default)]
    pub struct ActorMetrics;

    impl ActorMetrics {
        pub fn new() -> Self {
            Self
        }
        #[inline]
        pub fn record_turn(&self, _high_depth: usize, _low_depth: usize) {}
        #[inline]
        pub fn record_priority_choice(&self, _took_high: bool, _low_queued: usize) {}
        #[inline]
        pub fn record_hold(&self, _kind: CommandKind, _held: Duration) {}
    }
}

pub use imp::ActorMetrics;

/// Saturation point for a recorded hold, in microseconds (~2,000 years).
///
/// Exists so the packed `longest` field cannot have a pathological duration
/// overflow into the kind bits. A hold this long is not a measurement, it is a
/// hang — and the counter should stay readable rather than start reporting the
/// wrong command.
///
/// Kept out of the `metrics` cfg so the invariant test below runs in the default
/// build too: the packing is a property of the layout, and a build that does not
/// record is exactly the build where nobody would notice it break.
#[allow(dead_code)]
const MICROS_CEILING: u64 = (1u64 << 56) - 1;

/// One command kind's holds, as of the moment [`ActorMetrics::snapshot`] read it.
#[cfg(feature = "metrics")]
#[derive(Debug, Clone, PartialEq, Eq)]
#[non_exhaustive]
pub struct KindSnapshot {
    pub kind: CommandKind,
    /// Turns spent on this kind.
    pub turns: u64,
    /// Turns that exceeded [`crate::CHUNK_BUDGET`]. Always 0 for the kinds
    /// [`CommandKind::exempt_from_budget`] names — see there for why, and for
    /// the criterion that decides which those are.
    ///
    /// **Occurrences, not magnitude**: one per turn that exceeded, however far
    /// it exceeded by. A kind whose hold has doubled reports the same count and
    /// a different [`Self::longest`].
    pub over_budget: u64,
    pub mean: Duration,
    /// This kind's longest hold. Distinct from [`MetricsSnapshot::longest`],
    /// which names one command across all kinds and so tends to be permanently
    /// whichever kind is slowest overall.
    pub longest: Duration,
    /// Counts per [`BUCKET_BOUNDS_MICROS`], plus a final overflow bucket.
    ///
    /// Private behind [`Self::buckets`] since 0.12.8 (W4.2). A public array
    /// field publishes `BUCKET_COUNT` as part of the type's shape, so adding a
    /// bucket bound would break every caller that named the length — and the
    /// bounds are exactly the thing a latency histogram is likely to want to
    /// re-cut. The accessor returns a slice and the length becomes an
    /// observation rather than a signature. Python already did it this way.
    buckets: [u64; BUCKET_COUNT],
}

#[cfg(feature = "metrics")]
impl KindSnapshot {
    /// Counts per [`BUCKET_BOUNDS_MICROS`], plus a final overflow bucket.
    ///
    /// Pair it with `BUCKET_BOUNDS_MICROS` to label the axis rather than
    /// hard-coding the bounds; the slice is one longer than that constant,
    /// and the extra trailing element is the overflow bucket.
    pub fn buckets(&self) -> &[u64] {
        &self.buckets
    }
}

/// What the actor has done since the database was opened.
#[cfg(feature = "metrics")]
#[derive(Debug, Clone, PartialEq)]
#[non_exhaustive]
pub struct MetricsSnapshot {
    /// Commands executed, i.e. the sum of [`KindSnapshot::turns`]. The two agree
    /// by construction rather than by coincidence.
    pub turns: u64,
    /// Loop iterations that took a queue-depth reading. Always at least
    /// `turns + 1` on a live actor, because the reading is taken on the way in
    /// to a `select!` that has not resolved yet. This is the denominator of the
    /// two means below, and it is exposed so the difference is visible rather
    /// than looking like drift.
    pub depth_samples: u64,
    pub high_depth_mean: f64,
    pub high_depth_max: u64,
    pub low_depth_mean: f64,
    pub low_depth_max: u64,
    /// The longest hold since open and what caused it. `None` before the first
    /// turn, and — honestly — also when every turn so far took under a
    /// microsecond, which on this path does not happen.
    pub longest: Option<(CommandKind, Duration)>,
    /// Turns spent on high-priority work while low-priority work was already
    /// queued (0.12.10, W4.4, D-153).
    ///
    /// The actor's `select!` is `biased` and has **no floor**, so this is the
    /// measurement of a bound the design has always had and never observed.
    /// On its own it is not alarming: a busy database *should* prefer
    /// interactive writes, and this counter rising is that working. Read it
    /// beside [`Self::low_starved_run_max`], which is the number with teeth.
    pub low_starved_turns: u64,
    /// The longest unbroken run of the above — i.e. the most turns any single
    /// low-priority command has waited (0.12.10, W4.4, D-153).
    ///
    /// This is the one that answers "can low-priority work be starved". A large
    /// `low_starved_turns` spread over a long session says the tiers are doing
    /// their job; a large *run* says one specific chunk, rebuild or archive sat
    /// behind that many interactive writes in a row.
    ///
    /// # There is deliberately no forced-yield policy, and the reason changed
    /// (0.13.26, W10.4, [D-199])
    ///
    /// It used to be "adding one now would be fixing a bound nobody has
    /// observed being hit". That premise died twice. [D-153] hit the bound
    /// completely on a synthetic burst, and W10.4 then hit it on an ordinary
    /// one: **four closed-loop writers** — each awaiting its own write before
    /// issuing the next, which is what application code does — starve the low
    /// tier for essentially all of their writes
    /// (`examples/fairness_probe.rs`). The run is bounded by how long the
    /// caller keeps offering interactive work, not by concurrency and not by
    /// anything in this crate.
    ///
    /// **What replaced it is the floor's own price.** "After N starved turns,
    /// take one low-priority command" cannot choose *which* command — the low
    /// queue is an mpsc channel and its head is not inspectable — and at least
    /// one low-priority kind is exempt from [`crate::CHUNK_BUDGET`] **by
    /// contract**: an [`crate::Database::archive`] was measured at 3.3 s
    /// unwindowed on an 8,000-key backlog. So the floor would add an unbounded
    /// term to the interactive worst case in order to unblock work that is
    /// declared not to be latency-sensitive, which is the tier split running
    /// backwards.
    ///
    /// **The lever that does work belongs to the caller**: 1 ms of think time
    /// between a writer's writes takes four writers from ~78 to ~2. Which makes
    /// this field the instrument for a decision the caller owns rather than a
    /// defect report about the actor.
    ///
    /// [D-153]: ../docs/architecture/s13-decision-register.md#d-153
    /// [D-199]: ../docs/architecture/s13-decision-register.md#d-199
    pub low_starved_run_max: u64,
    pub kinds: Vec<KindSnapshot>,
}

#[cfg(feature = "metrics")]
impl MetricsSnapshot {
    /// Kinds that broke the budget, worst first. The one-line answer to "is the
    /// 3 ms bound holding?".
    pub fn budget_violations(&self) -> Vec<&KindSnapshot> {
        let mut v: Vec<_> = self.kinds.iter().filter(|k| k.over_budget > 0).collect();
        v.sort_by_key(|k| std::cmp::Reverse(k.over_budget));
        v
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn every_kind_indexes_to_its_own_slot() {
        for (i, &kind) in CommandKind::ALL.iter().enumerate() {
            assert_eq!(kind.index(), i, "{kind} is out of order in ALL");
        }
        assert_eq!(CommandKind::COUNT, CommandKind::ALL.len());
    }

    /// The budget is a bucket boundary, not a value inside one — so "fits in the
    /// budget" is a prefix sum and needs no interpolation.
    #[test]
    fn the_chunk_budget_is_exactly_a_bucket_boundary() {
        let budget = crate::CHUNK_BUDGET.as_micros() as u64;
        assert!(
            BUCKET_BOUNDS_MICROS.contains(&budget),
            "CHUNK_BUDGET is {budget} µs, which is not a bucket bound: \
             {BUCKET_BOUNDS_MICROS:?}"
        );
        assert_eq!(bucket_of(budget), bucket_of(budget - 1));
        assert_eq!(bucket_of(budget + 1), bucket_of(budget) + 1);
    }

    #[test]
    fn the_overflow_bucket_catches_everything_past_the_last_bound() {
        let last = *BUCKET_BOUNDS_MICROS.last().unwrap();
        assert_eq!(bucket_of(last), BUCKET_BOUNDS_MICROS.len() - 1);
        assert_eq!(bucket_of(last + 1), BUCKET_COUNT - 1);
        assert_eq!(bucket_of(u64::MAX), BUCKET_COUNT - 1);
    }

    /// The packing is the reason `longest` is one atomic: duration high, kind
    /// low, so a `fetch_max` on the word compares the duration.
    #[test]
    fn the_packing_leaves_room_for_both_fields() {
        assert!(
            (CommandKind::COUNT as u64) <= 0xFF,
            "the kind index must fit in the low 8 bits"
        );
        // The ceiling must survive being shifted up by the kind's width.
        assert_eq!(MICROS_CEILING.checked_shl(8), Some(MICROS_CEILING << 8));
        assert_eq!((MICROS_CEILING << 8) >> 8, MICROS_CEILING);
    }

    /// The two halves of a shadow rebuild land on opposite sides of the
    /// budget, and a forged hold is the only way to assert it (0.14.16, D-233).
    ///
    /// The integration suite can run a real rebuild and check that the
    /// violation list comes back empty; what it cannot do is make a *fill*
    /// chunk run long on demand. So the canary lives here, where the hold is an
    /// argument: the same over-budget duration recorded against each half must
    /// produce a violation for one and not the other.
    ///
    /// Without this, widening the exemption to cover both halves would leave
    /// every test in the crate green — `a_swap_over_budget_is_not_a_violation`
    /// included, since it asserts a zero that a broader exemption also
    /// produces. This is the assertion that says the zero means *healthy* and
    /// not *unwatched*.
    #[cfg(feature = "metrics")]
    #[test]
    fn a_long_fill_is_a_violation_and_a_long_swap_is_not() {
        let m = ActorMetrics::new();
        let over = crate::CHUNK_BUDGET + Duration::from_millis(44);

        m.record_hold(CommandKind::ShadowRebuild, over);
        m.record_hold(CommandKind::ShadowSwap, over);

        let snap = m.snapshot();
        let of = |kind: CommandKind| {
            snap.kinds
                .iter()
                .find(|k| k.kind == kind)
                .unwrap()
                .over_budget
        };

        assert_eq!(
            of(CommandKind::ShadowRebuild),
            1,
            "a fill chunk ran {over:?} against a {:?} budget and was not \
             counted. The fill half is the canary D-082 refused to exempt and \
             D-233 kept unexempted; if it stops counting, a regression on the \
             one path the chunked rebuild exists to keep short is invisible.",
            crate::CHUNK_BUDGET
        );
        assert_eq!(
            of(CommandKind::ShadowSwap),
            0,
            "the swap was counted as a violation. It exceeds by construction \
             on every healthy database, so counting it makes \
             `budget_violations()` nonzero forever (D-233)."
        );

        // And the magnitude survives the exemption, which is the half of the
        // argument that decided C over B: exempting removes the *occurrence*
        // from the violation list and touches nothing a reader consults to see
        // the hold grow.
        let longest = snap
            .kinds
            .iter()
            .find(|k| k.kind == CommandKind::ShadowSwap)
            .unwrap()
            .longest;
        assert_eq!(
            longest, over,
            "the swap's hold stopped being recorded when it stopped being \
             counted. `over_budget` counts occurrences; the histogram and \
             `longest` are where growth is visible, and an exemption must not \
             reach them."
        );
    }

    #[cfg(feature = "metrics")]
    #[test]
    fn the_longest_hold_names_the_command_that_caused_it() {
        let m = ActorMetrics::new();
        m.record_hold(CommandKind::AssertEdge, Duration::from_micros(500));
        m.record_hold(CommandKind::Archive, Duration::from_millis(40));
        m.record_hold(CommandKind::UpsertConcept, Duration::from_micros(900));

        let snap = m.snapshot();
        assert_eq!(
            snap.longest,
            Some((CommandKind::Archive, Duration::from_millis(40)))
        );
    }

    /// The regression the packing bug produced: a *short* hold of a
    /// later-declared kind must not outrank a long hold of an earlier one.
    ///
    /// The test above does not catch it, because `Archive` happens to be both
    /// the longest hold and a high enum index — which is exactly why the first
    /// version of the packing shipped past it. Here the two orderings disagree.
    #[cfg(feature = "metrics")]
    #[test]
    fn a_later_declared_kind_does_not_outrank_a_longer_hold() {
        let long = CommandKind::AssertEdge; // index 0
        let short = CommandKind::RebuildFts; // last index
        assert!(short.index() > long.index(), "the fixture needs the gap");

        let m = ActorMetrics::new();
        m.record_hold(long, Duration::from_millis(40));
        m.record_hold(short, Duration::from_micros(1));

        assert_eq!(
            m.snapshot().longest,
            Some((long, Duration::from_millis(40))),
            "the max is being taken over the kind index, not the duration"
        );
    }

    /// The three contractual exemptions must not show up as violations, or the
    /// violation count is noise on any database that archives.
    #[cfg(feature = "metrics")]
    #[test]
    fn an_exempt_kind_over_budget_is_not_a_violation() {
        let m = ActorMetrics::new();
        m.record_hold(CommandKind::Archive, Duration::from_millis(40));
        m.record_hold(CommandKind::AssertEdge, Duration::from_millis(40));

        let snap = m.snapshot();
        let violations = snap.budget_violations();
        assert_eq!(violations.len(), 1);
        assert_eq!(violations[0].kind, CommandKind::AssertEdge);
        assert_eq!(violations[0].over_budget, 1);

        // But the hold is still *recorded* — exempt means "not a violation",
        // not "not measured". A 40 ms archive is exactly what T1.1 exists to
        // shrink, and it cannot be shrunk if it is not counted.
        let archive = snap
            .kinds
            .iter()
            .find(|k| k.kind == CommandKind::Archive)
            .unwrap();
        assert_eq!(archive.turns, 1);
        assert_eq!(archive.mean, Duration::from_millis(40));
    }

    #[cfg(feature = "metrics")]
    #[test]
    fn queue_depth_is_a_mean_and_a_high_water_mark() {
        let m = ActorMetrics::new();
        m.record_turn(0, 4);
        m.record_turn(10, 0);

        let snap = m.snapshot();
        // No command ran, so `turns` is 0 while `depth_samples` is 2. The two
        // counters are different facts and this is the case that shows it.
        assert_eq!(snap.turns, 0);
        assert_eq!(snap.depth_samples, 2);
        assert_eq!(snap.high_depth_mean, 5.0);
        assert_eq!(snap.high_depth_max, 10);
        assert_eq!(snap.low_depth_mean, 2.0);
        assert_eq!(snap.low_depth_max, 4);
    }
}