prikk_store/

layout.rs

//! Repository layout paths and initialization.

use std::path::{Path, PathBuf};

use prikk_error::{PrikkError, Result};
use prikk_hash::{sha256, to_hex};
use prikk_object::{ObjectId, ObjectType, is_windows_reserved_name};

use crate::fsutil::{
    MutationRoot, create_new_file_required, ensure_directory_required, read_file_if_exists,
    read_file_required,
};

const REPO_DIR: &str = ".prikk";
const LEGACY_FORMAT_VERSION: &[u8] = b"1\n";
const LEGACY_FORMAT_2_VERSION: &[u8] = b"2\n";
const LEGACY_FORMAT_3_VERSION: &[u8] = b"3\n";
const LEGACY_FORMAT_4_VERSION: &[u8] = b"4\n";
const LEGACY_FORMAT_5_VERSION: &[u8] = b"5\n";
/// Visible at `pub(crate)` so `format_stability_gate.rs`'s pinning test can assert this byte form
/// and `CURRENT_FORMAT_VERSION_NUMERIC` against literal expected values side by side.
pub(crate) const CURRENT_FORMAT_VERSION: &[u8] = b"6\n";
/// Numeric companion to `CURRENT_FORMAT_VERSION`, for `format_stability_gate.rs`'s own range
/// arithmetic (RFC 114 §4, Gate B layer 1). Kept as an independent literal, not parsed from
/// `CURRENT_FORMAT_VERSION` at compile or test time -- see
/// `format_stability_gate::current_format_version_byte_and_numeric_forms_agree` for why.
#[cfg(test)]
pub(crate) const CURRENT_FORMAT_VERSION_NUMERIC: u32 = 6;

/// Repository format selected by the authoritative `.prikk/FORMAT` marker.
///
/// RFC 103 retired format 1; RFC 102 Stage 3 retired format 2 the same way, RFC 102 Stage 4 did the
/// same again for format 3, RFC 102 Stage 5 did it once more for format 4, and RFC 102 Stage 6 does it
/// again for format 5 -- rejected at open (`read_repository_format`), not merely unsupported for
/// mutation, no variant naming it here. **This bump was sought and decided by the owner explicitly**
/// (design-v1.md §14.7, 2026-08-15) and Stage 6 follows the same precedent (design-v1.md §15.6):
/// Stage 6 Step 1 adds a B slot and a generation log for each of the three compacting containers, new
/// names `durable_append`'s strictness makes unsafe to leave undetected in an older repository. The
/// single remaining variant is kept as an enum rather than collapsed away, per design-v1.md §12.1's
/// own note: `require_current_format`'s disk re-read is a real runtime check (RFC 103 Increment B was
/// abandoned specifically because of it), so the enum's *shape* still carries meaning and is not free
/// to simplify away.
///
/// **"Format 2"/"format 3"/"format 4"/"format 5"/"format 6" here name the on-disk repository layout**
/// (loose objects/refs vs. RFC 102's containers) **-- a different axis from DC-40's "format-2" wire
/// schema** (`block_state.rs`, `state_root.rs`, `format.rs`'s Block/Patch shape and Merkle rules),
/// which no RFC 102 stage touches and which keeps its own "format-2" name regardless of what this
/// enum's current variant is called (design §8: "format-2's rejection of the ahead-log state" is
/// explicitly unchanged).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum RepositoryFormat {
    /// Current format 6: RFC 102 Stage 6 Step 1's generation-aware index containers (ref pointer
    /// index, received-ref index, trust policy container), layered on every prior stage's container
    /// work, still writable under the unchanged DC-40 schema and state-root rules.
    CurrentV6,
}

/// One container's pre-allocated alternate slot (RFC's §3.2 compaction requirement: a fixed A/B pair
/// of names, never a rotated/new name). Object and ref-log containers keep `B` reserved-but-unused
/// forever, per design-v1.md §15.2 -- object compaction has no data model to target and the ref log
/// must never be compacted (DC-38/DC-69). The three genuine compaction targets (ref pointer index,
/// received-ref index, trust policy container -- design-v1.md §15.1) got their own `A`/`B` slots in
/// Stage 6 Step 1; `B` is written only once Stage 6 Step 2's compactor exists.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ContainerSlot {
    /// The slot every write targets until compaction (Stage 6 Step 2) exists.
    A,
    /// The alternate slot compaction publishes to. Unused by object/ref-log containers, which never
    /// compact; unused by the three Step-1-generation-aware containers until Step 2 lands.
    B,
}

impl ContainerSlot {
    /// Return a stable lower-case label used in the container's file name.
    #[must_use]
    pub const fn as_str(self) -> &'static str {
        match self {
            Self::A => "a",
            Self::B => "b",
        }
    }

    /// Return the other slot -- the compactor's own "which slot am I retiring into" question
    /// (design-v1.md §15.6/§4: compaction always writes the slot that is *not* currently live).
    #[must_use]
    pub const fn other(self) -> Self {
        match self {
            Self::A => Self::B,
            Self::B => Self::A,
        }
    }
}

/// One of the four containers RFC 102 Stage 6 Step 2 locks against concurrent writer/compactor races
/// (design-v1.md §15.8, ruled **wide** by the project owner over the developer's own narrower lean):
/// the three genuine compaction targets, plus the ref log, whose own tearing exposure predates RFC 102
/// and is not caused by compaction, but is fixed here because the exclusion machinery being built for
/// compaction closes it for free. `trust_key_container` is deliberately absent -- it never compacts,
/// and stays protected by the unchanged, repository-wide `ActiveLock` alone, the same as before this
/// stage.
///
/// `derive(Ord)` on a fieldless enum compares by declaration order, which **is** the one fixed total
/// lock order every multi-container acquisition sorts into (`lock::acquire_container_locks`,
/// design-v1.md §15.7's deadlock ruling) -- no call site can express an inverted order even by
/// accident, because sorting is structural, not a discipline to remember.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub enum LockableContainer {
    /// `refs/containers/pointer-index-{a,b}.container` -- Stage 6 Step 1's own generation-aware
    /// target.
    RefPointerIndex,
    /// `refs/containers/log-{a,b}.container` -- never compacted (DC-38/DC-69), but shares the ref
    /// pointer index's write path (`publication.rs`) and the same unserialized-appender exposure.
    RefLog,
    /// `refs/containers/received-index-{a,b}.container` -- Stage 6 Step 1's own generation-aware
    /// target, and the container whose investigation (`bundle.rs:270`'s `import_bundle`, no lock at
    /// all) is what surfaced this whole ruling.
    ReceivedIndex,
    /// `trust/policy-{a,b}.container` -- Stage 6 Step 1's own generation-aware target. Already
    /// incidentally protected today by `ActiveLock` (`trust.rs:88,144`); gains its own dedicated lock
    /// here anyway, per the owner's decision that the lock is container-scoped, not repository-wide
    /// (design-v1.md §15.7 decision 2) -- so a `prikk compact` run on this container never contends
    /// with unrelated `ActiveLock` holders (a `commit`, a `seal`) that never touch it.
    TrustPolicy,
}

impl LockableContainer {
    /// Every variant, in the declared total order -- the one enumeration `prikk unlock` (and anything
    /// else that needs to sweep every container lock) should use, rather than hand-rolling the list
    /// and risking it drift from the enum's own declaration.
    pub const ALL: [Self; 4] = [
        Self::RefPointerIndex,
        Self::RefLog,
        Self::ReceivedIndex,
        Self::TrustPolicy,
    ];
}

/// Repository layout paths.
#[derive(Debug, Clone)]
pub struct RepositoryLayout {
    root: PathBuf,
    prikk_dir: PathBuf,
    worktree_mutation: MutationRoot,
    repository_mutation: MutationRoot,
    format: RepositoryFormat,
}

impl PartialEq for RepositoryLayout {
    fn eq(&self, other: &Self) -> bool {
        self.root == other.root && self.prikk_dir == other.prikk_dir
    }
}

impl Eq for RepositoryLayout {}

impl RepositoryLayout {
    /// Create a layout for a working tree root.
    pub fn new(root: impl Into<PathBuf>) -> Result<Self> {
        let root = root.into();
        let prikk_dir = root.join(REPO_DIR);
        let worktree_mutation = MutationRoot::open(&root)?;
        let repository_mutation = worktree_mutation.open_root(Path::new(REPO_DIR))?;
        let format = read_repository_format(&repository_mutation)?;
        Ok(Self {
            root,
            prikk_dir,
            worktree_mutation,
            repository_mutation,
            format,
        })
    }

    /// Initialize a repository layout on disk.
    pub fn init(root: impl Into<PathBuf>) -> Result<Self> {
        let root = root.into();
        let prikk_dir = root.join(REPO_DIR);
        let worktree_mutation = MutationRoot::open(&root)?;
        let repository_mutation = worktree_mutation.ensure_root(Path::new(REPO_DIR))?;
        if let Some(version) = read_file_if_exists(&repository_mutation, Path::new("FORMAT"))? {
            if version != CURRENT_FORMAT_VERSION {
                // RFC 102 Stage 3, design-v1.md §12.1: this refusal inherited the format-2 -> 3 bump;
                // Stage 4 carried it forward for format-3 -> 4, Stage 5 did the same for format-4 ->
                // 5, and Stage 6 does it again for format-5 -> 6 (design-v1.md §15.6). Audited, not
                // just the constant swapped: unlike `read_repository_format`'s own rejection (reached
                // via `open`), this fires only on a redundant `init` against an already-initialized
                // repository of some other format, so it stays terse and points at `open` (any other
                // command) for the detailed migration message rather than duplicating it here.
                return Err(PrikkError::Integrity(
                    "refusing to initialize an existing non-format-6 Prikk repository (open it \
                     with any other command for a detailed unsupported-format message)"
                        .to_string(),
                ));
            }
        }
        let layout = Self {
            root,
            prikk_dir,
            worktree_mutation,
            repository_mutation,
            format: RepositoryFormat::CurrentV6,
        };
        for dir in layout.required_repository_directories()? {
            ensure_directory_required(layout.repository_mutation_root(), &dir)?;
        }
        // RFC 102 Stage 1: created at `init`, never later -- a missing file and an idempotent
        // re-`init` on an already-initialized repository must not clobber either.
        create_empty_file_once(&layout, &layout.worktree_unclean_shutdown_marker_path())?;
        create_empty_file_once(&layout, &layout.default_queue_wal_path())?;
        // RFC 102 Stage 5, design-v1.md §14.6: the active-WAL ref-ownership metadata, on the same
        // marker pattern as the worktree marker above -- created empty at `init`, set by truncate-then
        // -append, cleared by truncate-to-empty, never removed. Previously created lazily on the
        // empty-to-non-empty WAL transition (`active.rs::prepare_empty_active_ref_for_append`).
        create_empty_file_once(&layout, &layout.default_active_ref_name_path())?;
        // RFC 102 Stage 3, design-v1.md §2: every container name, both slots, plus the index and the
        // (currently unused) compaction generation log -- all allocated here, at `init`, and nowhere
        // else, for the life of the repository. This is the acceptance test itself (handoff §5
        // criterion 1): enumerate every one of these paths and confirm none of them is ever created
        // by any other code path.
        for object_type in persisted_object_types() {
            create_empty_file_once(
                &layout,
                &layout.container_slot_path(object_type, ContainerSlot::A),
            )?;
            create_empty_file_once(
                &layout,
                &layout.container_slot_path(object_type, ContainerSlot::B),
            )?;
        }
        create_empty_file_once(&layout, &layout.container_index_path())?;
        create_empty_file_once(&layout, &layout.container_generation_log_path())?;
        // RFC 102 Stage 4, Step 0 §13.2/§13.4: the shared ref-log container's both slots, plus the
        // separate ref-pointer-index container -- allocated here, at `init`, and nowhere else, the
        // same acceptance-criterion-1 discipline Stage 3 established.
        create_empty_file_once(
            &layout,
            &layout.ref_log_container_slot_path(ContainerSlot::A),
        )?;
        create_empty_file_once(
            &layout,
            &layout.ref_log_container_slot_path(ContainerSlot::B),
        )?;
        // RFC 102 Stage 6 Step 1, design-v1.md §15.6: the three genuine compaction targets each gain
        // an A/B pair and their own generation log here, allocated at `init` like every other name --
        // Step 1 itself never writes B or a generation record, so every one of these stays empty
        // until Step 2's compactor exists (handoff §2 criterion 2, "no behaviour change").
        create_empty_file_once(
            &layout,
            &layout.ref_pointer_index_slot_path(ContainerSlot::A),
        )?;
        create_empty_file_once(
            &layout,
            &layout.ref_pointer_index_slot_path(ContainerSlot::B),
        )?;
        create_empty_file_once(&layout, &layout.ref_pointer_index_generation_log_path())?;
        create_empty_file_once(&layout, &layout.received_index_slot_path(ContainerSlot::A))?;
        create_empty_file_once(&layout, &layout.received_index_slot_path(ContainerSlot::B))?;
        create_empty_file_once(&layout, &layout.received_index_generation_log_path())?;
        create_empty_file_once(&layout, &layout.trust_key_container_path())?;
        // DC-53 Stage 1: allocated at `init` like every other container name, for new repositories.
        // A repository initialized before this increment simply has no such file --
        // `author_key_index.rs` reads that identically to an empty container, so no format bump or
        // migration step is needed for existing repositories. True for writes too, not just reads:
        // `record_author_key_material` creates the container lazily on first write if it is still
        // absent (`author_key_index.rs::ensure_author_key_container_exists`), since
        // `append_file_required` -- unlike the read path -- requires the target to already exist.
        create_empty_file_once(&layout, &layout.author_key_container_path())?;
        create_empty_file_once(
            &layout,
            &layout.trust_policy_container_slot_path(ContainerSlot::A),
        )?;
        create_empty_file_once(
            &layout,
            &layout.trust_policy_container_slot_path(ContainerSlot::B),
        )?;
        create_empty_file_once(&layout, &layout.trust_policy_generation_log_path())?;
        // RFC 102 Stage 5, design-v1.md §14.2: written last, once every container/marker/WAL name
        // above is confirmed present. `FORMAT`'s presence is what certifies `init` completed --
        // written first (the old order), a crash between it and the containers left a repository
        // that read as a valid, empty format-4 repository with every container absent, and nothing
        // detected it (`status`/`verify`/`doctor` all exited 0 against a probe repository with all 16
        // container files deleted). Written last, an interrupted `init` leaves `FORMAT` absent, so a
        // re-`init` skips the mismatched-format guard above (it only fires when `FORMAT` already
        // exists) and re-enters this same body -- every `create_empty_file_once` call above is
        // idempotent, so the re-run completes whichever names are still missing and finishes by
        // writing `FORMAT`, exactly the "detectable and completable" property the reordering exists
        // to provide.
        //
        // RFC 102 Stage 5, design-v1.md §14.10: `create_new_file_required` (`create_exclusive`), not
        // `write_file_atomically` (`atomic_replace`) -- the same primitive `create_empty_file_once`
        // already uses for every other name above. For a name that does not yet exist,
        // `atomic_replace`'s rename-into-place is a new-directory-entry event, exactly the class this
        // RFC exists to eliminate; `create_exclusive` is one new-name event with no temp file and no
        // rename. Confirmed, not assumed: `create_exclusive` (`anchored/linux.rs`) syncs both the file
        // and the parent directory, the same durability `atomic_replace` provided. The `is_none()`
        // guard above still governs whether this runs at all -- unchanged -- but the create call
        // itself is now exclusive, so a genuine concurrent-`init` race now errors on `FORMAT` the same
        // way it already does on every other name `create_empty_file_once` allocates, rather than
        // FORMAT alone silently accepting whichever racer's rename landed last. Not a new failure
        // mode: it is FORMAT joining the behavior every other name in this function already has.
        if read_file_if_exists(layout.repository_mutation_root(), Path::new("FORMAT"))?.is_none() {
            create_new_file_required(
                layout.repository_mutation_root(),
                Path::new("FORMAT"),
                CURRENT_FORMAT_VERSION,
            )?;
        }
        Ok(layout)
    }

    /// Open an existing repository layout.
    pub fn open(root: impl Into<PathBuf>) -> Result<Self> {
        Self::new(root)
    }

    /// Return the repository format selected when this layout was opened.
    #[must_use]
    pub const fn format(&self) -> RepositoryFormat {
        self.format
    }

    pub(crate) fn validate_format(&self) -> Result<()> {
        let format = read_repository_format(self.repository_mutation_root())?;
        if format != self.format {
            return Err(PrikkError::UnsupportedFormatVersion(0));
        }
        Ok(())
    }

    /// Refuse ordinary repository/worktree mutation in legacy format 1.
    pub fn require_current_format(&self) -> Result<()> {
        self.validate_format()?;
        if self.format == RepositoryFormat::CurrentV6 {
            return Ok(());
        }
        Err(PrikkError::UnsupportedFormatVersion(1))
    }

    /// Return the working tree root.
    #[must_use]
    pub fn root(&self) -> &Path {
        &self.root
    }

    /// Return the `.prikk` directory.
    #[must_use]
    pub fn prikk_dir(&self) -> &Path {
        &self.prikk_dir
    }

    /// Return the repository format marker path.
    #[must_use]
    pub fn format_path(&self) -> PathBuf {
        self.prikk_dir.join("FORMAT")
    }

    /// Return the unclean-shutdown worktree marker path (RFC 102 Stage 1). Created empty at `init`;
    /// non-empty means worktree materialization was interrupted and commit-authoring must refuse to
    /// infer deletion from absence until the worktree is re-verified against its baseline. Always
    /// updated by append/truncate (`fsutil::append_file_required`/`truncate_file_empty_required`),
    /// never by `atomic_replace` -- RFC 102 §3's correction: `atomic_replace` renames over the
    /// destination unconditionally, which is a new-name event whose Windows durability is DC-87
    /// §3.4's still-open question, exactly the gap this marker exists to close.
    #[must_use]
    pub fn worktree_unclean_shutdown_marker_path(&self) -> PathBuf {
        self.prikk_dir.join("worktree.marker")
    }

    /// Return the object root directory.
    ///
    /// Dead-surface consolidation: no longer in `required_directories()` (RFC-format-3 repositories
    /// write containers, not loose object files, so a fresh repository never gets this tree). Kept
    /// `pub(crate)`, not removed, because `verify/objects.rs`'s `scan_loose_file_temp_debris` still
    /// reads it on every `verify` -- a dormant, format-3-only-relevant diagnostic (design-v1.md §12.3
    /// item 3) whose own retirement was explicitly reserved for an RFC-level act, not a side effect of
    /// this consolidation. `inspect_entry` tolerates the tree's absence, so a repository that never had
    /// it (or an old one that did) both verify cleanly.
    #[must_use]
    pub(crate) fn objects_dir(&self) -> PathBuf {
        self.prikk_dir.join("objects")
    }

    /// Return the active-session root directory.
    #[must_use]
    pub fn active_dir(&self) -> PathBuf {
        self.prikk_dir.join("active")
    }

    /// Return the default active-session directory.
    #[must_use]
    pub fn default_active_dir(&self) -> PathBuf {
        self.active_dir().join("default")
    }

    /// Return the default active WAL path.
    #[must_use]
    pub fn default_queue_wal_path(&self) -> PathBuf {
        self.default_active_dir().join("queue.wal")
    }

    /// Return the default active lock path.
    #[must_use]
    pub fn default_active_lock_path(&self) -> PathBuf {
        self.default_active_dir().join("active.lock")
    }

    /// Return the default active-session ref-name metadata path.
    #[must_use]
    pub fn default_active_ref_name_path(&self) -> PathBuf {
        self.default_active_dir().join("ref-name")
    }

    /// Return the ref root directory.
    #[must_use]
    pub fn refs_dir(&self) -> PathBuf {
        self.prikk_dir.join("refs")
    }

    /// Return the cache directory.
    #[must_use]
    pub fn cache_dir(&self) -> PathBuf {
        self.prikk_dir.join("cache")
    }

    /// Return the container root directory (RFC 102 Stage 3, design-v1.md §2).
    #[must_use]
    pub fn containers_dir(&self) -> PathBuf {
        self.prikk_dir.join("containers")
    }

    /// Return the container directory for a persisted object type.
    #[must_use]
    pub fn container_type_dir(&self, object_type: ObjectType) -> PathBuf {
        self.containers_dir()
            .join(object_type_directory_name(object_type))
    }

    /// Return one object type's container file for a given slot. **Every slot's name is allocated
    /// at `init`, including B, even though Stage 3 only ever writes A** -- compaction (Stage 6, not
    /// authorized) is what would ever target B; the RFC's §3.2 fixed-name-set requirement applies to
    /// the whole RFC, not per stage, so the name exists now regardless of when it is first used.
    #[must_use]
    pub fn container_slot_path(&self, object_type: ObjectType, slot: ContainerSlot) -> PathBuf {
        self.container_type_dir(object_type)
            .join(format!("{}.container", slot.as_str()))
    }

    /// Return the object index's container path. Single file, no A/B slot -- design-v1.md §4 /
    /// RFC 102's own §6.7 answer #2: the index's publication shape is plain append-only ("A/B" for an
    /// index reduces to "append-only wearing an A/B costume, not a second option" once forced through
    /// this codebase's real primitives), so unlike the six object-type containers it needs only one
    /// name.
    #[must_use]
    pub fn container_index_path(&self) -> PathBuf {
        self.containers_dir().join("index.container")
    }

    /// Return the small, fixed-name compaction generation log (design-v1.md §4: "compaction publishes
    /// by appending a generation record to a small fixed-name log; readers take the last complete
    /// generation record"). **Reserved, not used, by Stage 3** -- its name must still be allocated at
    /// `init` because compaction (Stage 6) is not authorized to create any name later. Absent any
    /// generation record (the only state Stage 3 ever produces), every container type's slot A is
    /// live by construction -- there is nothing for an empty log to disambiguate yet.
    #[must_use]
    pub fn container_generation_log_path(&self) -> PathBuf {
        self.containers_dir().join("generations.log")
    }

    /// Return the ref-container root directory (RFC 102 Stage 4). Kept under `refs/`, sibling to the
    /// now-vestigial `by-id/`/`logs/`/`tmp/`/`locks/` directories, rather than under the object
    /// `containers/` tree -- ref containers are not object containers and the two are never confused
    /// for the same purpose.
    #[must_use]
    pub fn refs_containers_dir(&self) -> PathBuf {
        self.refs_dir().join("containers")
    }

    /// Return the shared ref-log container file for a given slot (Step 0 §13.2: one container holds
    /// every ref's log records, forced by acceptance criterion 1 -- ref names do not exist at `init`,
    /// so a per-ref container is architecturally impossible). **Both slots allocated at `init`**,
    /// matching Stage 3's own A/B convention exactly (`container_slot_path`'s own doc comment) --
    /// Stage 4 only ever writes `A`.
    #[must_use]
    pub fn ref_log_container_slot_path(&self, slot: ContainerSlot) -> PathBuf {
        self.refs_containers_dir()
            .join(format!("log-{}.container", slot.as_str()))
    }

    /// Return the ref-pointer-index container path for a given slot (RFC 102 Stage 6 Step 1,
    /// design-v1.md §15.6: this is one of the three genuine compaction targets -- `ref_pointer_index`
    /// is last-entry-wins, and every ref update strands the previous entry, §15.1's own finding. `A`/`B`
    /// slots mirror `container_slot_path`'s own naming shape). Reads and writes resolve which slot is
    /// live through `generation.rs`'s resolver; Step 1 always resolves `A` because no generation record
    /// has ever been written -- see `ref_pointer_index_generation_log_path`.
    #[must_use]
    pub fn ref_pointer_index_slot_path(&self, slot: ContainerSlot) -> PathBuf {
        self.refs_containers_dir()
            .join(format!("pointer-index-{}.container", slot.as_str()))
    }

    /// Return the ref-pointer-index generation log path (RFC 102 Stage 6 Step 1, design-v1.md §15.6
    /// item 3/§4: readers take the last complete generation record; empty until Step 2's compactor
    /// ever writes one, at which point `A` stops being the unconditional answer). Its own name, not
    /// the pre-existing `container_generation_log_path()` -- that name was allocated for object-
    /// container compaction, which §15.1 establishes will never happen under the current content-
    /// addressed, no-GC data model, and a shared log across independently-compacting containers would
    /// let one corrupt record take down slot resolution for all of them at once (§15.6's own
    /// blast-radius reasoning).
    #[must_use]
    pub fn ref_pointer_index_generation_log_path(&self) -> PathBuf {
        self.refs_containers_dir()
            .join("pointer-index-generation.log")
    }

    /// Return the received-ref-index container path for a given slot (RFC 102 Stage 6 Step 1,
    /// design-v1.md §15.6: the second of the three genuine compaction targets, same last-entry-wins
    /// shape as the ref pointer index). Formerly `received_index_path`, single-name -- RFC 102 Stage 5,
    /// design-v1.md §14.1/Step 0 item 2's own reasoning for why `received.rs` belongs on the refs
    /// container+pointer-index pattern is unaffected by gaining a slot; only the publication shape
    /// (single-name vs. resolver-selected) changed.
    #[must_use]
    pub fn received_index_slot_path(&self, slot: ContainerSlot) -> PathBuf {
        self.refs_containers_dir()
            .join(format!("received-index-{}.container", slot.as_str()))
    }

    /// Return the received-ref-index generation log path. Its own name, for the same blast-radius
    /// reason `ref_pointer_index_generation_log_path` has its own.
    #[must_use]
    pub fn received_index_generation_log_path(&self) -> PathBuf {
        self.refs_containers_dir()
            .join("received-index-generation.log")
    }

    /// Return all required directories for layout creation.
    ///
    /// Dead-surface consolidation: `objects/` and its six type subdirectories, `refs/by-id/`,
    /// `refs/logs/`, and `quarantine/` are no longer created here -- nothing in a format-3 repository
    /// writes into any of them (containers replaced loose objects and per-ref files years ago). This is
    /// not a format change: `required_directories()` is consulted only here, at `init`; nothing
    /// validates it at `open`, so an existing repository keeps its now-unused directories harmlessly,
    /// and only a newly initialized one has fewer. `refs/tmp/` stays -- `refs/verify.rs`'s
    /// `candidate_issues` still scans it on every `verify`.
    #[must_use]
    pub fn required_directories(&self) -> Vec<PathBuf> {
        let mut dirs = Vec::new();
        dirs.push(self.containers_dir());
        for object_type in persisted_object_types() {
            dirs.push(self.container_type_dir(object_type));
        }
        dirs.push(self.active_dir());
        dirs.push(self.default_active_dir());
        dirs.push(self.refs_dir());
        dirs.push(self.refs_dir().join("locks"));
        dirs.push(self.refs_dir().join("tmp"));
        dirs.push(self.refs_containers_dir());
        dirs.push(self.trust_dir());
        dirs.push(self.cache_dir());
        dirs
    }

    pub(crate) fn repository_mutation_root(&self) -> &MutationRoot {
        &self.repository_mutation
    }

    pub(crate) fn worktree_mutation_root(&self) -> &MutationRoot {
        &self.worktree_mutation
    }

    pub(crate) fn repository_relative(&self, path: &Path) -> Result<PathBuf> {
        path.strip_prefix(&self.prikk_dir)
            .map(Path::to_path_buf)
            .map_err(|_| {
                PrikkError::Io("path is outside repository mutation authority".to_string())
            })
    }

    fn required_repository_directories(&self) -> Result<Vec<PathBuf>> {
        self.required_directories()
            .into_iter()
            .map(|path| self.repository_relative(&path))
            .collect()
    }

    /// Return the object directory for a persisted object type.
    ///
    /// `pub(crate)`, not public: its only reader is `verify/objects.rs`'s
    /// `scan_loose_file_temp_debris`, a dormant format-3-only diagnostic (see `objects_dir`'s own doc).
    #[must_use]
    pub(crate) fn object_type_dir(&self, object_type: ObjectType) -> PathBuf {
        self.objects_dir()
            .join(object_type_directory_name(object_type))
    }

    /// Return the storage path for a persisted object ID and type.
    ///
    /// No production caller by design, not by omission: this addresses the pre-container, format-1/2
    /// loose-object layout, which nothing in a live format-3 repository writes or reads. Kept `pub`
    /// specifically so `prikk-cli`'s format-transition fixtures (`tests/format_transition_support/`, a
    /// different crate) can construct legacy-shaped repositories without duplicating the private
    /// hex-prefixing scheme (`hex_prefix`, `layout.rs`) this method wraps.
    #[must_use]
    pub fn object_path(&self, object_type: ObjectType, id: ObjectId) -> PathBuf {
        let hex = id.to_hex();
        let prefix = hex_prefix(&hex);
        self.object_type_dir(object_type)
            .join(prefix)
            .join(format!("{hex}.pobj"))
    }

    /// Return the flat ref pointer path for a human-readable ref name.
    ///
    /// No production caller by design: this addresses the pre-container, format-1/2 flat-pointer-file
    /// layout containers replaced. Kept `pub` so `prikk-cli`'s format-transition fixtures (a different
    /// crate) can construct legacy-shaped repositories without reimplementing the private
    /// `ref_name_storage_key` hash this method wraps -- that helper is `pub(crate)`, unreachable from
    /// outside this crate.
    #[must_use]
    pub fn ref_pointer_path(&self, ref_name: &str) -> PathBuf {
        self.refs_dir()
            .join("by-id")
            .join(format!("{}.ref", ref_name_storage_key(ref_name)))
    }

    /// Return the ref log path for a human-readable ref name.
    ///
    /// No production caller by design, for the same reason as `ref_pointer_path`: a legacy-format path
    /// builder kept `pub` for `prikk-cli`'s cross-crate format-transition fixtures.
    #[must_use]
    pub fn ref_log_path(&self, ref_name: &str) -> PathBuf {
        self.refs_dir()
            .join("logs")
            .join(format!("{}.log", ref_name_storage_key(ref_name)))
    }

    /// Return the ref lock path for a human-readable ref name.
    #[must_use]
    pub fn ref_lock_path(&self, ref_name: &str) -> PathBuf {
        self.refs_dir()
            .join("locks")
            .join(format!("{}.lock", ref_name_storage_key(ref_name)))
    }

    /// Return the ref temporary candidate path for a human-readable ref name.
    ///
    /// No production writer since Stage 4 removed the candidate-write-then-promote mechanism, but this
    /// is not dead the way `ref_pointer_path`/`ref_log_path` are: `refs/tmp/` itself stays in
    /// `required_directories()` because `refs/verify.rs`'s `candidate_issues` scans it on every
    /// `verify`, and this accessor is what several of that scan's own regression tests
    /// (`refs/tests/publication_recovery/candidate_cleanup.rs`) use to construct debris inside it.
    #[must_use]
    pub fn ref_tmp_path(&self, ref_name: &str) -> PathBuf {
        self.refs_dir()
            .join("tmp")
            .join(format!("{}.tmp", ref_name_storage_key(ref_name)))
    }

    /// Return the publication trust-store directory.
    #[must_use]
    pub fn trust_dir(&self) -> PathBuf {
        self.prikk_dir.join("trust")
    }

    /// Return the trust key-material container path (RFC 102 Stage 5, design-v1.md §14/§14.9).
    /// Replaces the one-file-per-key-id `trust/keys/maintainer/*.pub` directory entirely -- format 5
    /// rejects every repository old enough to have one, so no repository this code can open ever
    /// contains that directory's contents (§14.9 §3's own reasoning, applied here as it was to
    /// `refs/received/`, not Stage 4's "keep, dead" precedent).
    #[must_use]
    pub fn trust_key_container_path(&self) -> PathBuf {
        self.trust_dir().join("keys.container")
    }

    /// Return the AUTHOR key-material container path (DC-53 Stage 1,
    /// `.git-exclude/reviewed/DC-53-stage-1-report-ruling-v1.md` §5). Its own container, not a
    /// third role folded into `trust_key_container_path` -- that one is MAINTAINER key material
    /// with a policy layered over it; this one is material only, populated by the authoring path
    /// rather than an adoption command (`author_key_index.rs`'s own module doc). A repository
    /// initialized before this container existed has no such file, which `author_key_index.rs`
    /// reads identically to an empty one, not a structural defect.
    #[must_use]
    pub fn author_key_container_path(&self) -> PathBuf {
        self.trust_dir().join("author-keys.container")
    }

    /// Return the trust policy container path for a given slot (RFC 102 Stage 5, design-v1.md
    /// §14/§14.9, gaining a slot in Stage 6 Step 1, design-v1.md §15.6 -- the third of the three
    /// genuine compaction targets: one complete snapshot appended per `add`/`remove`, every earlier
    /// snapshot dead, §15.1's own finding). Each append is a **complete snapshot** of the adopted key
    /// id list, not an incremental log entry -- see `trust_index.rs`'s own module doc for why that is
    /// what makes revocation representable without a tombstone record; that property is unaffected by
    /// gaining a slot.
    #[must_use]
    pub fn trust_policy_container_slot_path(&self, slot: ContainerSlot) -> PathBuf {
        self.trust_dir()
            .join(format!("policy-{}.container", slot.as_str()))
    }

    /// Return the trust-policy generation log path. Its own name, for the same blast-radius reason
    /// `ref_pointer_index_generation_log_path` has its own -- and distinct from `trust_key_container_
    /// path`, which is **not** one of the three compacting containers and gains no slot: TOFU history
    /// must persist across removal (`trust.rs:77`), which compacting the key container would break.
    #[must_use]
    pub fn trust_policy_generation_log_path(&self) -> PathBuf {
        self.trust_dir().join("policy-generation.log")
    }

    /// Return the lock file path for one of Stage 6 Step 2's four `LockableContainer`s
    /// (design-v1.md §15.8). Ephemeral, like every other lock file in this codebase
    /// (`ActiveLock`/`RefLock`): created on acquire, removed on release, never pre-allocated at
    /// `init` -- criterion 2's "every name created at `init`" obligation is about durability-bearing
    /// container names, not transient mutual-exclusion markers, and `ActiveLock`/`RefLock` already
    /// establish that a lock file is exempt from it.
    #[must_use]
    pub fn lockable_container_lock_path(&self, container: LockableContainer) -> PathBuf {
        match container {
            LockableContainer::RefPointerIndex => {
                self.refs_containers_dir().join("pointer-index.lock")
            }
            LockableContainer::RefLog => self.refs_containers_dir().join("log.lock"),
            LockableContainer::ReceivedIndex => {
                self.refs_containers_dir().join("received-index.lock")
            }
            LockableContainer::TrustPolicy => self.trust_dir().join("policy.lock"),
        }
    }
}

/// Validate a maintainer key id's storage safety: ASCII alphanumeric/`-`/`_` only, and not a
/// Windows-reserved device stem. Split out from the retired `maintainer_trust_key_path` (which paired
/// this check with building a per-key-id file path that no longer exists under the container model) --
/// the validation itself is unchanged and still required before a key id is accepted.
pub(crate) fn validate_maintainer_key_id_storage_safety(key_id: &str) -> Result<()> {
    if key_id.is_empty()
        || !key_id
            .bytes()
            .all(|byte| byte.is_ascii_alphanumeric() || byte == b'-' || byte == b'_')
    {
        return Err(PrikkError::InvalidName(
            "maintainer key id is not storage-safe".to_string(),
        ));
    }
    // DC-72: the allowlist above is character-shape only and does not exclude Windows-reserved
    // device stems (`CON`, `PRN`, ...) — `CON` is all ASCII-alphanumeric and would otherwise
    // pass. Checked regardless of host OS, matching `RepoPath`'s equivalent rule.
    if is_windows_reserved_name(key_id) {
        return Err(PrikkError::InvalidName(format!(
            "maintainer key id is a Windows reserved device name: {key_id}"
        )));
    }
    Ok(())
}

/// Create `path` empty if it does not already exist. Idempotent, so a retried or re-run `init`
/// against an already-initialized repository never clobbers it -- the same rule RFC 102 Stage 1
/// established for the worktree marker and the active WAL, now shared by every `init`-time file this
/// layout creates.
fn create_empty_file_once(layout: &RepositoryLayout, path: &Path) -> Result<()> {
    let relative = layout.repository_relative(path)?;
    if read_file_if_exists(layout.repository_mutation_root(), &relative)?.is_none() {
        create_new_file_required(layout.repository_mutation_root(), &relative, &[])?;
    }
    Ok(())
}

#[cfg(test)]
mod tests;

fn read_repository_format(root: &MutationRoot) -> Result<RepositoryFormat> {
    let version = read_file_required(root, Path::new("FORMAT"))?;
    match version.as_slice() {
        // RFC 114 §5.3 (owner decision): formats 1-5 are not supported. A real released version can
        // still read this one -- 0.19.0 and earlier, confirmed against the release record -- so the
        // message must not claim otherwise; it states withdrawal, not absence.
        LEGACY_FORMAT_VERSION => Err(PrikkError::Integrity(
            "this repository uses format 1, which prikk no longer supports (this version \
             requires format 6). format-1 support was removed after 0.19.0; migration from \
             format 1 is not supported."
                .to_string(),
        )),
        // RFC 102 Stage 3, design-v1.md §12.1 (owner decision): bump to format 3, reject format-2 at
        // open, no dual-layout bridge. Format 2 was current through the 0.19.0 release (the same
        // release format-1's own rejection message above already points at), so it is the last
        // release able to open a format-2 repository -- established from the release record
        // (`CHANGELOG.md`, `git tag`), not guessed.
        // RFC 114 §5.3: same as format 1's arm -- 0.19.0 genuinely reads format 2 (confirmed against
        // the release record), so this states withdrawal, not absence of a reader.
        LEGACY_FORMAT_2_VERSION => Err(PrikkError::Integrity(
            "this repository uses format 2, which prikk no longer supports (this version \
             requires format 6). format-2 support was removed after 0.19.0; migration from \
             format 2 is not supported."
                .to_string(),
        )),
        // RFC 102 Stage 4: bump to format 4, reject format-3 at open, no dual-layout bridge --
        // applying Stage 3's own already-ruled policy (design-v1.md §12.1), not a fresh decision (see
        // `RepositoryFormat`'s own doc comment). Unlike format 2, format 3 was never itself the
        // subject of a tagged release (Stage 3 was still pending three-platform CI when Stage 4
        // began) -- no specific "removed after X.Y.Z" version is named here, since none can be
        // verified from the release record yet; naming one would be guessing, which this project's
        // own discipline for these messages does not do.
        // RFC 114 §5.3: unlike formats 1-2, no released version ever shipped format 3 -- the message
        // says so plainly rather than offering a step the product cannot honour.
        LEGACY_FORMAT_3_VERSION => Err(PrikkError::Integrity(
            "this repository uses format 3, which prikk no longer supports (this version \
             requires format 6). format 3 was never used in a released prikk version; there is \
             no supported migration path."
                .to_string(),
        )),
        // RFC 102 Stage 5, design-v1.md §14.7 (owner decision): bump to format 5, reject format-4 at
        // open, no dual-layout bridge -- format 3's precedent, not format 2's. No release was ever
        // tagged at format 4 either (Stage 4 merged and Stage 5 began before any tag), verified against
        // the release record (`CHANGELOG.md`, `git tag`) rather than assumed from format 3's own
        // no-version precedent -- naming one anyway would be guessing, which this project's own
        // discipline for these messages does not do.
        // RFC 114 §5.3: same as format 3's arm -- format 4 never shipped in a released version either.
        LEGACY_FORMAT_4_VERSION => Err(PrikkError::Integrity(
            "this repository uses format 4, which prikk no longer supports (this version \
             requires format 6). format 4 was never used in a released prikk version; there is \
             no supported migration path."
                .to_string(),
        )),
        // RFC 102 Stage 6, design-v1.md §15.6 (owner decision): bump to format 6, reject format-5 at
        // open, no dual-layout bridge -- the same precedent again. No release was ever tagged at
        // format 5 either (still 0.19.0, format 2, verified against `git tag` fresh rather than
        // assumed from the format-4 arm's own finding), so no version is named here for the same
        // reason.
        // RFC 114 §5.3: same as formats 3-4's arms -- format 5 never shipped in a released version.
        LEGACY_FORMAT_5_VERSION => Err(PrikkError::Integrity(
            "this repository uses format 5, which prikk no longer supports (this version \
             requires format 6). format 5 was never used in a released prikk version; there is \
             no supported migration path."
                .to_string(),
        )),
        CURRENT_FORMAT_VERSION => Ok(RepositoryFormat::CurrentV6),
        _ => Err(PrikkError::UnsupportedFormatVersion(0)),
    }
}

/// Return persisted object types. RefUpdate is log-inline in v1 and is intentionally absent.
#[must_use]
pub fn persisted_object_types() -> [ObjectType; 7] {
    [
        ObjectType::Patch,
        ObjectType::Block,
        ObjectType::RefState,
        ObjectType::Tag,
        ObjectType::Attestation,
        ObjectType::Blob,
        ObjectType::RecognitionClaim,
    ]
}

/// Return a stable directory name for an object type.
#[must_use]
pub fn object_type_directory_name(object_type: ObjectType) -> &'static str {
    match object_type {
        ObjectType::Patch => "patch",
        ObjectType::Block => "block",
        ObjectType::RefState => "ref-state",
        ObjectType::Tag => "tag",
        ObjectType::Attestation => "attestation",
        ObjectType::Blob => "blob",
        ObjectType::RecognitionClaim => "recognition-claim",
        ObjectType::RefUpdate => "ref-update-inline-only",
        // New FDD-03 §3 types. Full storage-layout placement (`objects/genesis/`,
        // `cache/block-summary/`, `refs/recovery/`) is reconciled in the FDD-02
        // layout phase; these names keep the mapper exhaustive without creating
        // directories yet.
        ObjectType::BlockSummaryCache => "block-summary-cache-rebuildable",
        ObjectType::RecoveryNote => "recovery-note-inline-only",
        ObjectType::ProjectGenesis => "genesis",
    }
}

fn hex_prefix(hex: &str) -> String {
    hex.chars().take(2).collect()
}

pub(crate) fn ref_name_storage_key(ref_name: &str) -> String {
    to_hex(&ref_name_key_bytes(ref_name))
}

/// The raw 32-byte form of [`ref_name_storage_key`] (RFC 102 Stage 4, Step 0 §13.4 / design-v1.md
/// §13.4's ruling): a fixed-width key already used to name every ref pointer/log file today, reused
/// as the ref-pointer-index's own key rather than inventing a second one -- the "new key shape"
/// objection Step 0 raised dissolved specifically because this already existed.
#[must_use]
pub(crate) fn ref_name_key_bytes(ref_name: &str) -> [u8; 32] {
    sha256(ref_name.as_bytes())
}