diff --git a/gix-ref/Cargo.toml b/gix-ref/Cargo.toml
index 0ef687cccf0..c2c3dde9a69 100644
--- a/gix-ref/Cargo.toml
+++ b/gix-ref/Cargo.toml
@@ -26,7 +26,7 @@ serde = ["dep:serde", "gix-hash/serde", "gix-actor/serde", "gix-object/serde"]
 parallel = ["gix-features/parallel"]
 
 [dependencies]
-gix-features = { version = "^0.48.0", path = "../gix-features", features = ["walkdir"] }
+gix-features = { version = "^0.48.0", path = "../gix-features", features = ["walkdir", "once_cell"] }
 gix-fs = { version = "^0.21.1", path = "../gix-fs" }
 gix-path = { version = "^0.12.0", path = "../gix-path" }
 gix-hash = { version = "^0.25.0", path = "../gix-hash" }
diff --git a/gix-ref/src/store/packed/buffer.rs b/gix-ref/src/store/packed/buffer.rs
index 9c1a757c7fc..b2839fdd62e 100644
--- a/gix-ref/src/store/packed/buffer.rs
+++ b/gix-ref/src/store/packed/buffer.rs
@@ -67,6 +67,8 @@ pub mod open {
                 data: backing,
                 path,
                 object_hash,
+                lookup_count: std::sync::atomic::AtomicUsize::new(0),
+                name_index: gix_features::threading::OnceCell::new(),
             })
         }
 
diff --git a/gix-ref/src/store/packed/find.rs b/gix-ref/src/store/packed/find.rs
index 915e725029a..1e12c594f21 100644
--- a/gix-ref/src/store/packed/find.rs
+++ b/gix-ref/src/store/packed/find.rs
@@ -1,7 +1,16 @@
+use std::sync::atomic::Ordering;
+
 use gix_object::bstr::{BStr, BString, ByteSlice};
 
 use crate::{FullNameRef, PartialNameRef, store_impl::packed};
 
+/// Number of `try_find_full_name` calls served by binary search before
+/// the next call eagerly builds a name → offset index for the rest of the
+/// buffer's lifetime. The threshold trades a small one-time build cost
+/// against `log₂(n)`-per-call binary searches; below it, single-shot
+/// lookups (typical CLI commands) pay nothing extra.
+pub(super) const INDEX_BUILD_AFTER_LOOKUPS: usize = 8;
+
 /// packed-refs specific functionality
 impl packed::Buffer {
     /// Find a reference with the given `name` and return it.
@@ -37,6 +46,23 @@ impl packed::Buffer {
     }
 
     pub(crate) fn try_find_full_name(&self, name: &FullNameRef) -> Result<Option<packed::Reference<'_>>, Error> {
+        // Fast path: a built index turns the lookup into a single HashMap
+        // probe (plus an `decode::reference` re-parse at the matched offset).
+        if let Some(index) = self.name_index.get() {
+            return self.lookup_via_index(name, index);
+        }
+        // Cold path: count this lookup. After enough binary searches we
+        // amortize the index build cost, so build it now and serve this call
+        // from the index too.
+        let prev_lookups = self.lookup_count.fetch_add(1, Ordering::Relaxed);
+        if prev_lookups + 1 >= INDEX_BUILD_AFTER_LOOKUPS {
+            let index = self.name_index.get_or_init(|| self.build_name_index());
+            return self.lookup_via_index(name, index);
+        }
+        self.try_find_full_name_via_binary_search(name)
+    }
+
+    fn try_find_full_name_via_binary_search(&self, name: &FullNameRef) -> Result<Option<packed::Reference<'_>>, Error> {
         match self.binary_search_by(name.as_bstr()) {
             Ok(line_start) => {
                 let mut input = &self.as_ref()[line_start..];
@@ -54,6 +80,66 @@ impl packed::Buffer {
         }
     }
 
+    fn lookup_via_index(
+        &self,
+        name: &FullNameRef,
+        index: &packed::NameIndex,
+    ) -> Result<Option<packed::Reference<'_>>, Error> {
+        match index.by_name.get(name.as_bstr()) {
+            Some(&offset) => {
+                let mut input = &self.as_ref()[offset..];
+                Ok(Some(
+                    packed::decode::reference(&mut input, self.object_hash).map_err(|_| Error::Parse)?,
+                ))
+            }
+            None => {
+                if index.encountered_parse_failure {
+                    Err(Error::Parse)
+                } else {
+                    Ok(None)
+                }
+            }
+        }
+    }
+
+    /// Walk the entire buffer once, collecting `(name → record-start offset)`
+    /// pairs for every well-formed record. Malformed lines are skipped to
+    /// the next newline but flip [`packed::NameIndex::encountered_parse_failure`]
+    /// so the lookup path still reports them via [`Error::Parse`] on a miss.
+    fn build_name_index(&self) -> packed::NameIndex {
+        let body = self.as_ref();
+        let body_start = body.as_ptr() as usize;
+        let mut by_name = std::collections::HashMap::with_capacity(estimate_record_count(body));
+        let mut encountered_parse_failure = false;
+        let mut cursor = body;
+        loop {
+            if cursor.is_empty() {
+                break;
+            }
+            let record_offset = cursor.as_ptr() as usize - body_start;
+            let mut after = cursor;
+            match packed::decode::reference(&mut after, self.object_hash) {
+                Ok(r) => {
+                    by_name.insert(r.name.as_bstr().to_owned(), record_offset);
+                    cursor = after;
+                }
+                Err(_) => {
+                    encountered_parse_failure = true;
+                    // Skip to the next line so a single malformed record
+                    // doesn't prevent us from indexing the rest.
+                    cursor = match cursor.iter().position(|&b| b == b'\n') {
+                        Some(pos) => &cursor[pos + 1..],
+                        None => &[],
+                    };
+                }
+            }
+        }
+        packed::NameIndex {
+            by_name,
+            encountered_parse_failure,
+        }
+    }
+
     /// Find a reference with the given `name` and return it.
     pub fn find<'a, Name, E>(&self, name: Name) -> Result<packed::Reference<'_>, existing::Error>
     where
@@ -137,6 +223,17 @@ pub mod existing {
     }
 }
 
+/// A rough upper-bound estimate on the number of records in a packed-refs
+/// buffer, used only to preallocate the HashMap built by
+/// [`packed::Buffer::build_name_index`]. The shortest possible record is
+/// `<hash><space><name>\n` which is at least `hex_len + 4` bytes (1-char name).
+/// Overestimating is fine — the HashMap just grows naturally.
+fn estimate_record_count(body: &[u8]) -> usize {
+    // Conservative average of ~50 bytes/record gives a sensible initial
+    // capacity without scanning the file.
+    body.len() / 50
+}
+
 pub(crate) fn transform_full_name_for_lookup(name: &FullNameRef) -> Option<&FullNameRef> {
     match name.category_and_short_name() {
         Some((c, sn)) => {
diff --git a/gix-ref/src/store/packed/mod.rs b/gix-ref/src/store/packed/mod.rs
index 6da4d7b7e37..bd040cda865 100644
--- a/gix-ref/src/store/packed/mod.rs
+++ b/gix-ref/src/store/packed/mod.rs
@@ -1,5 +1,6 @@
-use std::path::PathBuf;
+use std::{collections::HashMap, path::PathBuf, sync::atomic::AtomicUsize};
 
+use gix_features::threading::OnceCell;
 use gix_hash::ObjectId;
 use gix_object::bstr::{BStr, BString};
 use memmap2::Mmap;
@@ -26,6 +27,35 @@ pub struct Buffer {
     offset: usize,
     /// The path from which we were loaded
     path: PathBuf,
+    /// Number of [`Self::try_find_full_name`] calls served by the binary
+    /// search so far. After it crosses [`find::INDEX_BUILD_AFTER_LOOKUPS`]
+    /// the next lookup eagerly builds [`name_index`] for O(1) lookups, so
+    /// single-shot callers (typical CLI commands) don't pay the build cost.
+    pub(crate) lookup_count: AtomicUsize,
+    /// Lazily populated map from ref name to the offset of its record in
+    /// [`AsRef::as_ref`]. Built once per buffer snapshot and consulted on
+    /// every subsequent lookup, bypassing the per-call binary search.
+    pub(crate) name_index: OnceCell<NameIndex>,
+}
+
+/// A name → record-offset index for fast `try_find_full_name` lookups.
+///
+/// Stored alongside [`Buffer`] inside a [`OnceCell`] so that once built it
+/// is shared by every clone of the snapshot. Records that don't parse are
+/// still detected: their presence flips [`Self::encountered_parse_failure`]
+/// so a miss against a corrupt packed-refs file surfaces as
+/// [`find::Error::Parse`] rather than `Ok(None)`, matching the binary-search
+/// path's behavior.
+#[derive(Debug)]
+pub(crate) struct NameIndex {
+    /// Maps a ref's full name to the byte offset of its record's first
+    /// byte in [`Buffer::as_ref`] (i.e. the hash byte, not the line break
+    /// before it).
+    pub(crate) by_name: HashMap<BString, usize>,
+    /// True if any record in the file failed to parse while the index was
+    /// being built. A missed lookup is reported as `Error::Parse` when this
+    /// is set, mirroring the binary-search path's parse-failure flag.
+    pub(crate) encountered_parse_failure: bool,
 }
 
 struct Edit {
diff --git a/gix-ref/tests/refs/packed/find.rs b/gix-ref/tests/refs/packed/find.rs
index 48bd7493ccf..d94751adbef 100644
--- a/gix-ref/tests/refs/packed/find.rs
+++ b/gix-ref/tests/refs/packed/find.rs
@@ -190,6 +190,194 @@ bogus refs/tags/gix-actor-v0.1.0
     Ok(())
 }
 
+/// Build a packed-refs body with `count` valid records whose names are
+/// `refs/heads/auto-NNNN`. The body is sorted, satisfying the "sorted"
+/// header so `packed::Buffer::open` keeps the original byte layout.
+fn synthetic_packed_refs(count: usize) -> Vec<u8> {
+    let mut out: Vec<u8> = b"# pack-refs with: peeled fully-peeled sorted\n".to_vec();
+    for i in 0..count {
+        let line = format!("{:040x} refs/heads/auto-{i:04}\n", i as u128 + 1);
+        out.extend_from_slice(line.as_bytes());
+    }
+    out
+}
+
+/// Mixed corpus: `count` valid records plus a malformed line spliced in
+/// the middle. Returned names are sorted so the file remains valid.
+fn synthetic_packed_refs_with_corruption(count: usize) -> Vec<u8> {
+    let mut out: Vec<u8> = b"# pack-refs with: peeled fully-peeled sorted\n".to_vec();
+    for i in 0..count {
+        let line = format!("{:040x} refs/heads/auto-{i:04}\n", i as u128 + 1);
+        out.extend_from_slice(line.as_bytes());
+        if i == count / 2 {
+            // A line with the wrong shape (no space after the would-be hash) —
+            // `decode::reference` will reject it while leaving the surrounding
+            // records intact.
+            out.extend_from_slice(b"bogus refs/heads/auto-malformed\n");
+        }
+    }
+    out
+}
+
+#[test]
+fn many_lookups_keep_returning_correct_results_across_the_index_threshold() -> crate::Result {
+    // Cross the lazy-index build threshold and verify every lookup — both
+    // the binary-search ones below it and the index-served ones above it —
+    // returns the same record we'd iterate.
+    let body = synthetic_packed_refs(64);
+    let (_keep, path) = write_packed_refs_with(&body)?;
+    let buf = packed::Buffer::open(path, 1024, HASH_KIND)?;
+
+    let names: Vec<_> = (0..64).map(|i| format!("refs/heads/auto-{i:04}")).collect();
+    for name in &names {
+        let found = buf.try_find(name.as_str())?.expect("ref is present");
+        assert_eq!(found.name.as_bstr(), name.as_str(), "name round-trips");
+    }
+    // A miss against a clean index reports `Ok(None)`, not `Err(Parse)`.
+    assert!(
+        buf.try_find("refs/heads/does-not-exist")?.is_none(),
+        "missing refs return Ok(None) when no malformed records are present"
+    );
+    Ok(())
+}
+
+#[test]
+fn index_path_surfaces_parse_failures_on_miss() -> crate::Result {
+    // Build a body large enough that `try_find` will trip the index-build
+    // threshold, then verify that a miss surfaces `Error::Parse` because a
+    // malformed record was encountered while building the index — matching
+    // the binary-search path's `encountered_parse_failure` semantics.
+    let body = synthetic_packed_refs_with_corruption(32);
+    let (_keep, path) = write_packed_refs_with(&body)?;
+    let buf = packed::Buffer::open(path, 1024, HASH_KIND)?;
+
+    // First do enough valid lookups to push past `INDEX_BUILD_AFTER_LOOKUPS`
+    // and force the index to be built.
+    for i in 0..16 {
+        let name = format!("refs/heads/auto-{i:04}");
+        let found = buf.try_find(name.as_str())?.expect("ref exists");
+        assert_eq!(found.name.as_bstr(), name.as_str());
+    }
+
+    // Now ask for a name that doesn't exist. Because the malformed record
+    // is in the file, the index records `encountered_parse_failure=true`,
+    // and a miss surfaces as a parse error rather than `Ok(None)`.
+    assert_eq!(
+        buf.try_find("refs/heads/never-there")
+            .expect_err("miss against a corrupt file must surface as Error::Parse")
+            .to_string(),
+        "The reference could not be parsed",
+        "index lookup matches the binary-search path's behavior on miss when corruption exists",
+    );
+    Ok(())
+}
+
+/// Profiling helper: simulate a wide-refs fetch's `update_refs` loop and
+/// break down where the time goes. Run with:
+///
+/// ```text
+/// cargo test -p gix-ref --release --features sha1 --test refs \
+///     -- --ignored --nocapture loose_stat_overhead_profile
+/// ```
+///
+/// Reports total time for two strategies against a ~150k-packed-ref repo:
+///   A. `store.try_find(name)` — the current `try_find_reference` path,
+///      which `stat`s the loose-ref file before falling through to packed.
+///   B. Pre-built `HashSet` of loose names → bypass the stat when the name
+///      isn't in it, then `packed.try_find(name)` directly.
+///
+/// The delta (A − B) is the loose-stat overhead that a hypothetical
+/// `file::Store`-level loose-name cache (or a caller-side enumeration like
+/// the original #2605) would reclaim on top of this PR's packed-buffer
+/// index. Useful for deciding whether such a follow-up is worth its
+/// complexity, since the answer is workload-dependent (warm vs cold cache,
+/// loose-ref count, filesystem characteristics).
+#[test]
+#[ignore = "profiling only; expensive and times vary across machines"]
+fn loose_stat_overhead_profile() -> crate::Result {
+    use std::collections::HashSet;
+    use std::time::Instant;
+
+    let store = store_at("make_repository_with_lots_of_packed_refs.sh")?;
+    let packed = store.open_packed_buffer()?.expect("packed-refs present");
+
+    // Collect all 150k packed ref names — these are what fetch's update_refs
+    // would iterate over.
+    let collect_start = Instant::now();
+    let packed_names: Vec<_> = packed
+        .iter()?
+        .filter_map(Result::ok)
+        .map(|r| r.name.to_owned())
+        .collect();
+    eprintln!(
+        "Collected {} packed ref names in {:?}",
+        packed_names.len(),
+        collect_start.elapsed()
+    );
+
+    // Strategy A — current behavior, with loose-stat per call.
+    let warmup = Instant::now();
+    for name in packed_names.iter().take(1000) {
+        let _ = store.try_find(name.as_ref())?;
+    }
+    eprintln!("Warmup (1000 lookups): {:?}", warmup.elapsed());
+
+    let a_start = Instant::now();
+    let mut a_found = 0usize;
+    for name in &packed_names {
+        if store.try_find(name.as_ref())?.is_some() {
+            a_found += 1;
+        }
+    }
+    let a_elapsed = a_start.elapsed();
+    eprintln!(
+        "Strategy A (store.try_find — per-call loose stat): {:?}, found {} of {} ({:.1} µs/lookup)",
+        a_elapsed,
+        a_found,
+        packed_names.len(),
+        a_elapsed.as_secs_f64() * 1_000_000.0 / packed_names.len() as f64,
+    );
+
+    // Strategy B — simulated loose-index: pre-enumerate loose names, skip stat
+    // when name isn't present.
+    let loose_set_start = Instant::now();
+    let loose_set: HashSet<_> = store.loose_iter()?.filter_map(Result::ok).map(|r| r.name).collect();
+    eprintln!(
+        "Built loose-name set ({} entries) in {:?}",
+        loose_set.len(),
+        loose_set_start.elapsed()
+    );
+
+    let b_start = Instant::now();
+    let mut b_found = 0usize;
+    for name in &packed_names {
+        if loose_set.contains(name) {
+            // Take slow path for shadowed entries.
+            if store.try_find(name.as_ref())?.is_some() {
+                b_found += 1;
+            }
+        } else if packed.try_find(name.as_ref())?.is_some() {
+            b_found += 1;
+        }
+    }
+    let b_elapsed = b_start.elapsed();
+    eprintln!(
+        "Strategy B (loose-set short-circuit + packed direct): {:?}, found {} of {} ({:.1} µs/lookup)",
+        b_elapsed,
+        b_found,
+        packed_names.len(),
+        b_elapsed.as_secs_f64() * 1_000_000.0 / packed_names.len() as f64,
+    );
+
+    eprintln!(
+        "Loose-stat overhead reclaimable by an in-Store loose-name cache: {:?} ({:.1}% of A)",
+        a_elapsed.saturating_sub(b_elapsed),
+        (a_elapsed.saturating_sub(b_elapsed).as_secs_f64() / a_elapsed.as_secs_f64()) * 100.0,
+    );
+    assert_eq!(a_found, b_found, "both strategies find the same number of refs");
+    Ok(())
+}
+
 #[test]
 fn find_speed() -> crate::Result {
     let store = store_at("make_repository_with_lots_of_packed_refs.sh")?;