perf(images): probe raster headers instead of decoding whole payloads, memoize repo icon validation (#18421)

* perf(images): measure raster headers from a probe and memoize repo icon validation

`getRepos()` re-sanitizes every repo on every call, and an uploaded/file repo
icon costs a full base64 decode of its data URI each time. Three fixes:

- `writeQuartet` destructured a mutable array, which sends V8 through the
  iterator protocol once per four input characters; index reads plus a length
  counter produce identical bytes.
- `decodeBase64Prefix` decoded the whole payload despite only the first bytes
  being needed. `exceedsRasterImagePreviewLimits` now probes 64 bytes and
  widens x16 until the header measures, and only re-runs the original
  full-payload decode when the verdict would suppress a preview.
- `sanitizeRepoIcon`'s src validation is memoized per icon source with a
  bounded FIFO map, reusing the `memoizeTitleClassification` idiom (now a
  shared `memoizeByStringKey`).

* perf(images): key icon-validation memo on the persisted icon object

Replaces the per-source 64-entry FIFO string-key memo with a WeakMap keyed on
the persisted repoIcon object that hydrateRepo already receives, storing
{src, source, supported} and re-checking both fields on a hit.

Retention becomes zero by construction (entries die with state.repos[i].repoIcon),
so there is no cap to evict live icons and no dead icon strings held after a repo
or icon is replaced. The identity re-check makes an in-place mutation unable to
serve a stale verdict. Drops bounded-string-key-memo.ts and reverts the collateral
terminal-title-classification-memo refactor.

src/shared/raster-image-base64-preview.test.ts

@@ -0,0 +1,345 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+import { decodeBase64Prefix, exceedsRasterImagePreviewLimits } from './raster-image-base64-preview'
+import type * as RasterImageDimensionsModule from './raster-image-dimensions'
+import { readRasterImageDimensions } from './raster-image-dimensions'
+import {
+  isKnownRasterImageMimeType,
+  isRasterImagePreviewDimensions,
+  RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES
+} from './raster-image-preview-limits'
+
+/** The pre-change verdict: one full-payload decode, then one dimension read. */
+function unprobedExceeds(content: string, mimeType: string | undefined): boolean {
+  if (!isKnownRasterImageMimeType(mimeType)) {
+    return false
+  }
+  const prefix = referenceDecodeBase64Prefix(content, RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES)
+  if (!prefix) {
+    return false
+  }
+  const dimensions = readRasterImageDimensions(prefix)
+  return dimensions !== null && !isRasterImagePreviewDimensions(dimensions)
+}
+
+// Why a module mock: the byte length handed to the dimension reader is the direct measure of how
+// much of the payload the preview check decoded, and it is the only observable difference between
+// the early-stopping probe and the full-payload decode it replaces.
+const { dimensionReadLengths } = vi.hoisted(() => ({ dimensionReadLengths: [] as number[] }))
+
+vi.mock('./raster-image-dimensions', async (importOriginal) => {
+  const actual = await importOriginal<typeof RasterImageDimensionsModule>()
+  return {
+    ...actual,
+    readRasterImageDimensions: (bytes: Uint8Array) => {
+      dimensionReadLengths.push(bytes.byteLength)
+      return actual.readRasterImageDimensions(bytes)
+    }
+  }
+})
+
+// ── Reference decoder: the pre-change implementation, verbatim ──────────────────────────────────
+const BASE64_PADDING = -2
+const INVALID_BASE64 = -1
+
+function base64Value(code: number): number {
+  if (code >= 65 && code <= 90) {
+    return code - 65
+  }
+  if (code >= 97 && code <= 122) {
+    return code - 71
+  }
+  if (code >= 48 && code <= 57) {
+    return code + 4
+  }
+  if (code === 43) {
+    return 62
+  }
+  if (code === 47) {
+    return 63
+  }
+  if (code === 61) {
+    return BASE64_PADDING
+  }
+  return INVALID_BASE64
+}
+
+function isWhitespace(code: number): boolean {
+  return code === 9 || code === 10 || code === 12 || code === 13 || code === 32
+}
+
+function referenceWriteQuartet(
+  output: Uint8Array,
+  offset: number,
+  quartet: readonly number[]
+): { bytesWritten: number; padded: boolean } | null {
+  const [a, b, c, d] = quartet
+  if (a === undefined || b === undefined || a < 0 || b < 0) {
+    return null
+  }
+  if (c === BASE64_PADDING) {
+    if (d !== BASE64_PADDING) {
+      return null
+    }
+    if (offset < output.length) {
+      output[offset] = (a << 2) | (b >> 4)
+    }
+    return { bytesWritten: Math.min(1, output.length - offset), padded: true }
+  }
+  if (c === undefined || c < 0) {
+    return null
+  }
+  if (offset < output.length) {
+    output[offset] = (a << 2) | (b >> 4)
+  }
+  if (offset + 1 < output.length) {
+    output[offset + 1] = ((b & 15) << 4) | (c >> 2)
+  }
+  if (d === BASE64_PADDING) {
+    return { bytesWritten: Math.min(2, output.length - offset), padded: true }
+  }
+  if (d === undefined || d < 0) {
+    return null
+  }
+  if (offset + 2 < output.length) {
+    output[offset + 2] = ((c & 3) << 6) | d
+  }
+  return { bytesWritten: Math.min(3, output.length - offset), padded: false }
+}
+
+function referenceDecodeBase64Prefix(content: string, maxBytes: number): Uint8Array | null {
+  const capacity = Math.min(maxBytes, Math.ceil(content.length / 4) * 3)
+  const output = new Uint8Array(capacity)
+  const quartet: number[] = []
+  let outputLength = 0
+  let padded = false
+
+  for (let index = 0; index < content.length && outputLength < capacity; index += 1) {
+    const code = content.charCodeAt(index)
+    if (isWhitespace(code)) {
+      continue
+    }
+    if (padded) {
+      return null
+    }
+    const value = base64Value(code)
+    if (value === INVALID_BASE64) {
+      return null
+    }
+    quartet.push(value)
+    if (quartet.length !== 4) {
+      continue
+    }
+    const decoded = referenceWriteQuartet(output, outputLength, quartet)
+    if (!decoded) {
+      return null
+    }
+    outputLength += decoded.bytesWritten
+    padded = decoded.padded
+    quartet.length = 0
+  }
+
+  if (!padded && outputLength < capacity && quartet.length > 0) {
+    if (quartet.length === 1 || quartet.includes(BASE64_PADDING)) {
+      return null
+    }
+    while (quartet.length < 4) {
+      quartet.push(BASE64_PADDING)
+    }
+    const decoded = referenceWriteQuartet(output, outputLength, quartet)
+    if (!decoded) {
+      return null
+    }
+    outputLength += decoded.bytesWritten
+  }
+  return output.subarray(0, outputLength)
+}
+
+// ── Fixtures ───────────────────────────────────────────────────────────────────────────────────
+function pngBytes(totalBytes: number, width: number, height: number): Buffer {
+  const bytes = Buffer.alloc(Math.max(totalBytes, 24))
+  Buffer.from([137, 80, 78, 71, 13, 10, 26, 10]).copy(bytes)
+  bytes.writeUInt32BE(13, 8)
+  bytes.write('IHDR', 12, 'ascii')
+  bytes.writeUInt32BE(width, 16)
+  bytes.writeUInt32BE(height, 20)
+  for (let index = 24; index < bytes.length; index += 1) {
+    bytes[index] = (index * 31 + 7) & 0xff
+  }
+  return bytes
+}
+
+/** SOI, `metadataBytes` of APP2 padding (real cameras chain many segments), then SOF0. */
+function jpegBytes(
+  metadataBytes: number,
+  width: number,
+  height: number,
+  trailingBytes = 4096
+): Buffer {
+  const parts: Buffer[] = [Buffer.from([0xff, 0xd8])]
+  for (let written = 0; written < metadataBytes;) {
+    const size = Math.min(65_533, metadataBytes - written)
+    const header = Buffer.alloc(4)
+    header.writeUInt16BE(0xffe2)
+    header.writeUInt16BE(size + 2, 2)
+    parts.push(header, Buffer.alloc(size))
+    written += size
+  }
+  const sof = Buffer.alloc(11)
+  sof.writeUInt16BE(0xffc0)
+  sof.writeUInt16BE(8, 2)
+  sof[4] = 8
+  sof.writeUInt16BE(height, 5)
+  sof.writeUInt16BE(width, 7)
+  parts.push(sof, Buffer.alloc(trailingBytes))
+  return Buffer.concat(parts)
+}
+
+function gifBytes(width: number, height: number): Buffer {
+  const gif = Buffer.alloc(64)
+  gif.write('GIF89a', 0, 'ascii')
+  gif.writeUInt16LE(width, 6)
+  gif.writeUInt16LE(height, 8)
+  return gif
+}
+
+function webpBytes(width: number, height: number): Buffer {
+  const webp = Buffer.alloc(64)
+  webp.write('RIFF', 0, 'ascii')
+  webp.writeUInt32LE(50, 4)
+  webp.write('WEBP', 8, 'ascii')
+  webp.write('VP8X', 12, 'ascii')
+  webp.writeUInt32LE(10, 16)
+  webp.writeUIntLE(width - 1, 24, 3)
+  webp.writeUIntLE(height - 1, 27, 3)
+  return webp
+}
+
+const DECODE_FIXTURES: { label: string; content: string }[] = [
+  { label: 'png', content: pngBytes(24, 512, 512).toString('base64') },
+  { label: 'png padded once', content: pngBytes(26, 512, 512).toString('base64') },
+  { label: 'png padded twice', content: pngBytes(25, 512, 512).toString('base64') },
+  { label: 'png 70 KiB', content: pngBytes(70_000, 512, 512).toString('base64') },
+  { label: 'jpeg', content: jpegBytes(0, 640, 480).toString('base64') },
+  { label: 'jpeg 70 KiB exif', content: jpegBytes(70_000, 4000, 3000).toString('base64') },
+  { label: 'gif', content: gifBytes(320, 240).toString('base64') },
+  { label: 'webp', content: webpBytes(800, 600).toString('base64') },
+  { label: 'empty', content: '' },
+  { label: 'one character', content: 'A' },
+  { label: 'two characters', content: 'AB' },
+  { label: 'three characters', content: 'ABC' },
+  { label: 'invalid character', content: 'AB*D' },
+  { label: 'invalid tail', content: `${pngBytes(24, 4, 4).toString('base64')}!!!` },
+  { label: 'padding mid-payload', content: 'AAAA=AAA' },
+  { label: 'lone padding in tail', content: 'AAAAAB=' },
+  { label: 'single padding', content: 'AAAAAA==' },
+  { label: 'double padding', content: 'AAAAAAA=' },
+  { label: 'stray padding after padding', content: 'AAAA====' },
+  {
+    label: 'line-wrapped png',
+    content: pngBytes(70_000, 512, 512)
+      .toString('base64')
+      .replace(/(.{76})/g, '$1\r\n')
+  },
+  { label: 'leading and trailing whitespace', content: `\n\t ${'AAAA'} \r\n` },
+  { label: 'truncated png header', content: pngBytes(24, 512, 512).toString('base64').slice(0, 18) }
+]
+
+const DECODE_CAPS = [
+  0,
+  1,
+  2,
+  3,
+  4,
+  23,
+  24,
+  25,
+  63,
+  64,
+  65,
+  1024,
+  RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES
+]
+
+describe('decodeBase64Prefix', () => {
+  it('decodes byte-for-byte identically to the reference implementation', () => {
+    for (const { label, content } of DECODE_FIXTURES) {
+      for (const maxBytes of DECODE_CAPS) {
+        const expected = referenceDecodeBase64Prefix(content, maxBytes)
+        const actual = decodeBase64Prefix(content, maxBytes)
+        const detail = `${label} @ maxBytes=${maxBytes}`
+        if (expected === null) {
+          expect(actual, detail).toBeNull()
+          continue
+        }
+        expect(actual, detail).not.toBeNull()
+        expect(Array.from(actual!), detail).toEqual(Array.from(expected))
+      }
+    }
+  })
+
+  it('stops at the byte cap instead of decoding the whole payload', () => {
+    const content = pngBytes(70_000, 512, 512).toString('base64')
+    expect(decodeBase64Prefix(content, 32)?.byteLength).toBe(32)
+  })
+})
+
+describe('exceedsRasterImagePreviewLimits', () => {
+  beforeEach(() => {
+    dimensionReadLengths.length = 0
+  })
+
+  it('measures a large image from its first bytes, not its last', () => {
+    // Regression guard: before the probe this decoded all 5 MiB before reading 24 bytes of IHDR.
+    const content = pngBytes(5 * 1024 * 1024, 512, 512).toString('base64')
+    expect(exceedsRasterImagePreviewLimits(content, 'image/png')).toBe(false)
+    expect(dimensionReadLengths).toEqual([64])
+  })
+
+  it('widens the probe until a JPEG SOF past its metadata is reachable', () => {
+    const content = jpegBytes(70_000, 4000, 3000, 3_000_000).toString('base64')
+    expect(exceedsRasterImagePreviewLimits(content, 'image/jpeg')).toBe(false)
+    expect(dimensionReadLengths).toEqual([64, 1024, 16_384, 262_144])
+    // Far below the ~3 MiB the payload decodes to, and below the 8 MiB fallback cap.
+    expect(dimensionReadLengths.at(-1)!).toBeLessThan(RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES)
+  })
+
+  it('re-reads the whole payload before suppressing an over-limit image', () => {
+    const content = pngBytes(70_000, 32_769, 1).toString('base64')
+    expect(exceedsRasterImagePreviewLimits(content, 'image/png')).toBe(true)
+    // The header answers at 64 bytes, but a suppression verdict is only taken from the same
+    // full-payload decode the unprobed implementation used, so invalid base64 past the header
+    // still demotes the answer to "could not measure".
+    expect(dimensionReadLengths).toEqual([64, 70_000])
+  })
+
+  it('keeps rendering an over-limit header whose payload is not valid base64', () => {
+    const content = `${pngBytes(70_000, 32_769, 1).toString('base64')}!!!`
+    expect(exceedsRasterImagePreviewLimits(content, 'image/png')).toBe(false)
+  })
+
+  it('agrees with the unprobed implementation on every fixture and mime type', () => {
+    const mimeTypes = [
+      'image/png',
+      'image/jpeg',
+      'image/gif',
+      'image/webp',
+      'image/svg+xml',
+      undefined
+    ]
+    const fixtures = [
+      ...DECODE_FIXTURES,
+      { label: 'over-limit png', content: pngBytes(24, 32_769, 1).toString('base64') },
+      { label: 'over-limit pixels png', content: pngBytes(24, 8192, 8192).toString('base64') },
+      { label: 'over-limit gif', content: gifBytes(65_535, 65_535).toString('base64') },
+      { label: 'over-limit jpeg', content: jpegBytes(70_000, 40_000, 40_000).toString('base64') },
+      { label: 'over-limit webp', content: webpBytes(40_000, 40_000).toString('base64') }
+    ]
+    for (const { label, content } of fixtures) {
+      for (const mimeType of mimeTypes) {
+        expect(exceedsRasterImagePreviewLimits(content, mimeType), `${label} / ${mimeType}`).toBe(
+          unprobedExceeds(content, mimeType)
+        )
+      }
+    }
+  })
+})

src/shared/raster-image-base64-preview.ts

@@ -34,13 +34,20 @@ function isWhitespace(code: number): boolean {
   return code === 9 || code === 10 || code === 12 || code === 13 || code === 32
 }
 
+/** `quartetLength` under 4 is a final short group; the missing slots decode as `=` padding. */
 function writeQuartet(
   output: Uint8Array,
   offset: number,
-  quartet: readonly number[]
+  quartet: readonly number[],
+  quartetLength: number
 ): { bytesWritten: number; padded: boolean } | null {
-  const [a, b, c, d] = quartet
-  if (a === undefined || b === undefined || a < 0 || b < 0) {
+  // Index reads, not `const [a, b, c, d] = quartet`: destructuring an array runs the iterator
+  // protocol (Symbol.iterator plus four `.next()` calls) once per four input characters.
+  const a = quartet[0]
+  const b = quartet[1]
+  const c = quartetLength > 2 ? quartet[2] : BASE64_PADDING
+  const d = quartetLength > 3 ? quartet[3] : BASE64_PADDING
+  if (quartetLength < 2 || a === undefined || b === undefined || a < 0 || b < 0) {
     return null
   }
   if (c === BASE64_PADDING) {
@@ -73,10 +80,13 @@ function writeQuartet(
   return { bytesWritten: Math.min(3, output.length - offset), padded: false }
 }
 
-function decodeBase64Prefix(content: string, maxBytes: number): Uint8Array | null {
+/** Exported so the decode can be compared byte-for-byte against a reference implementation. */
+export function decodeBase64Prefix(content: string, maxBytes: number): Uint8Array | null {
   const capacity = Math.min(maxBytes, Math.ceil(content.length / 4) * 3)
   const output = new Uint8Array(capacity)
-  const quartet: number[] = []
+  // Fixed four slots plus a counter, never resized: `quartet.length = 0` deoptimizes the array.
+  const quartet = [0, 0, 0, 0]
+  let quartetLength = 0
   let outputLength = 0
   let padded = false
 
@@ -92,27 +102,27 @@ function decodeBase64Prefix(content: string, maxBytes: number): Uint8Array | nul
     if (value === INVALID_BASE64) {
       return null
     }
-    quartet.push(value)
-    if (quartet.length !== 4) {
+    quartet[quartetLength] = value
+    quartetLength += 1
+    if (quartetLength !== 4) {
       continue
     }
-    const decoded = writeQuartet(output, outputLength, quartet)
+    const decoded = writeQuartet(output, outputLength, quartet, 4)
     if (!decoded) {
       return null
     }
     outputLength += decoded.bytesWritten
     padded = decoded.padded
-    quartet.length = 0
+    quartetLength = 0
   }
 
-  if (!padded && outputLength < capacity && quartet.length > 0) {
-    if (quartet.length === 1 || quartet.includes(BASE64_PADDING)) {
-      return null
+  if (!padded && outputLength < capacity && quartetLength > 0) {
+    for (let index = 0; index < quartetLength; index += 1) {
+      if (quartet[index] === BASE64_PADDING) {
+        return null
+      }
     }
-    while (quartet.length < 4) {
-      quartet.push(BASE64_PADDING)
-    }
-    const decoded = writeQuartet(output, outputLength, quartet)
+    const decoded = writeQuartet(output, outputLength, quartet, quartetLength)
     if (!decoded) {
       return null
     }
@@ -121,6 +131,13 @@ function decodeBase64Prefix(content: string, maxBytes: number): Uint8Array | nul
   return output.subarray(0, outputLength)
 }
 
+// First probe: past every fixed-offset header (PNG 24, GIF 10, WebP 30, BMP 26) and a JFIF-only
+// JPEG's SOF, so an icon or screenshot is measured from its first bytes instead of its last.
+const RASTER_IMAGE_HEADER_PROBE_BYTES = 64
+// Growth per miss. JPEG SOF sits past however much EXIF/ICC/MPF the camera wrote, so the probe
+// widens geometrically: total decoded stays within ~1.07x of the bytes the header actually needed.
+const RASTER_IMAGE_HEADER_PROBE_GROWTH = 16
+
 /**
  * Whether the encoded dimensions are known to exceed the preview limits.
  *
@@ -135,10 +152,34 @@ export function exceedsRasterImagePreviewLimits(
   if (!isKnownRasterImageMimeType(mimeType)) {
     return false
   }
-  const prefix = decodeBase64Prefix(content, RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES)
-  if (!prefix) {
-    return false
+  let probeBytes = RASTER_IMAGE_HEADER_PROBE_BYTES
+  for (;;) {
+    const prefix = decodeBase64Prefix(content, probeBytes)
+    // A short probe only ever fails where the whole payload would: it walks a strict prefix of the
+    // same characters through the same state machine.
+    if (!prefix) {
+      return false
+    }
+    // Shorter than asked for means the payload ran out, so a wider probe cannot add bytes.
+    const exhausted =
+      prefix.byteLength < probeBytes || probeBytes >= RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES
+    const dimensions = readRasterImageDimensions(prefix)
+    if (dimensions !== null) {
+      const withinLimits = isRasterImagePreviewDimensions(dimensions)
+      if (withinLimits || exhausted) {
+        return !withinLimits
+      }
+      // About to suppress: redo the decode over the whole payload so the verdict stays the one the
+      // full read gives, including its rejection of base64 that turns invalid past the header.
+      probeBytes = RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES
+      continue
+    }
+    if (exhausted) {
+      return false
+    }
+    probeBytes = Math.min(
+      probeBytes * RASTER_IMAGE_HEADER_PROBE_GROWTH,
+      RASTER_IMAGE_PREVIEW_HEADER_MAX_BYTES
+    )
   }
-  const dimensions = readRasterImageDimensions(prefix)
-  return dimensions !== null && !isRasterImagePreviewDimensions(dimensions)
 }

src/shared/repo-icon.test.ts

@@ -1,6 +1,22 @@
-import { describe, expect, it } from 'vitest'
+import { describe, expect, it, vi } from 'vitest'
+import type * as ImageDataUriModule from './image-data-uri'
 import { githubAvatarIcon, githubAvatarSlug, sanitizeRepoIcon } from './repo-icon'
 
+// Why a module mock: `validateRasterImageDataUri` is the leaf that base64-decodes an inline icon's
+// header, so counting its invocations is the direct measure of what re-hydrating a repo costs.
+const { dataUriValidations } = vi.hoisted(() => ({ dataUriValidations: { count: 0 } }))
+
+vi.mock('./image-data-uri', async (importOriginal) => {
+  const actual = await importOriginal<typeof ImageDataUriModule>()
+  return {
+    ...actual,
+    validateRasterImageDataUri: (dataUri: string) => {
+      dataUriValidations.count += 1
+      return actual.validateRasterImageDataUri(dataUri)
+    }
+  }
+})
+
 const PNG_1X1_BASE64 =
   'iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8/x8AAwMCAO+/p9sAAAAASUVORK5CYII='
 const WEBP_1X1_BASE64 = 'UklGRhoAAABXRUJQVlA4IA4AAAAwAQCdASoBAAEAAQIlSkwAAA=='
@@ -188,3 +204,84 @@ describe('githubAvatarSlug', () => {
     expect(githubAvatarSlug(null, undefined)).toBeNull()
   })
 })
+
+describe('repo icon source validation memo', () => {
+  const HYDRATIONS = 25
+
+  function uploadIcon(width: number): { type: 'image'; src: string; source: 'upload' } {
+    return { type: 'image', src: `data:image/png;base64,${pngBase64(width, 1)}`, source: 'upload' }
+  }
+
+  it('validates each distinct icon src once across repeated hydrations', () => {
+    const icons = [uploadIcon(2), uploadIcon(3), uploadIcon(4)]
+    // Warm the memo the way the first hydration would, then measure steady state.
+    for (const icon of icons) {
+      sanitizeRepoIcon(icon)
+    }
+    dataUriValidations.count = 0
+
+    for (let hydration = 0; hydration < HYDRATIONS; hydration += 1) {
+      for (const icon of icons) {
+        expect(sanitizeRepoIcon(icon)).toEqual(icon)
+      }
+    }
+
+    // Unmemoized this is HYDRATIONS x icons full base64 header decodes; memoized an unchanged
+    // persisted icon costs nothing.
+    expect(dataUriValidations.count).toBe(0)
+  })
+
+  it('re-validates as soon as the src changes', () => {
+    dataUriValidations.count = 0
+    expect(sanitizeRepoIcon(uploadIcon(11))).toEqual(uploadIcon(11))
+    expect(sanitizeRepoIcon(uploadIcon(12))).toEqual(uploadIcon(12))
+    expect(dataUriValidations.count).toBe(2)
+  })
+
+  it('keeps the verdict specific to the icon source', () => {
+    const src = `data:image/webp;base64,${WEBP_1X1_BASE64}`
+    expect(sanitizeRepoIcon({ type: 'image', src, source: 'file' })).toEqual({
+      type: 'image',
+      src,
+      source: 'file'
+    })
+    // WebP is a `file` icon only; sharing one cache across sources would accept it as an upload.
+    expect(sanitizeRepoIcon({ type: 'image', src, source: 'upload' })).toBeUndefined()
+  })
+
+  // Guard for the removed cap: the memo hangs off the persisted icon object, so it holds a verdict
+  // for every live icon no matter how many there are. A fixed-size map would evict the earliest
+  // entries here and re-decode them on the next hydration.
+  it('keeps a verdict for every live icon, however many repos have one', () => {
+    const LIVE_ICONS = 200
+    const icons = Array.from({ length: LIVE_ICONS }, (_, index) => uploadIcon(1000 + index))
+    for (const icon of icons) {
+      sanitizeRepoIcon(icon)
+    }
+    dataUriValidations.count = 0
+
+    for (const icon of icons) {
+      expect(sanitizeRepoIcon(icon)).toEqual(icon)
+    }
+
+    expect(dataUriValidations.count).toBe(0)
+  })
+
+  // Guard for the hazard object keying introduces: the stored src/source are re-checked on a hit,
+  // so a persisted icon edited in place can never be served its previous verdict.
+  it('re-validates an icon object whose src or source is mutated in place', () => {
+    const icon = { type: 'image', src: `data:image/png;base64,${pngBase64(7, 1)}`, source: 'file' }
+    expect(sanitizeRepoIcon(icon)).toEqual(icon)
+
+    icon.src = `data:image/png;base64,${pngBase64(8, 1)}`
+    dataUriValidations.count = 0
+    expect(sanitizeRepoIcon(icon)).toEqual(icon)
+    expect(dataUriValidations.count).toBe(1)
+
+    // WebP is a `file` icon but not an `upload` icon, so the same object must flip verdicts.
+    icon.src = `data:image/webp;base64,${WEBP_1X1_BASE64}`
+    expect(sanitizeRepoIcon(icon)).toEqual(icon)
+    icon.source = 'upload'
+    expect(sanitizeRepoIcon(icon)).toBeUndefined()
+  })
+})

src/shared/repo-icon.ts

@@ -79,7 +79,7 @@ function normalizeGitHubAvatarHost(rawHost?: string): string {
   }
 }
 
-function isSupportedImageSrc(src: string, source: RepoIconImageSource): boolean {
+function computeIsSupportedImageSrc(src: string, source: RepoIconImageSource): boolean {
   if (source === 'upload') {
     return (
       /^data:image\/png;base64,[A-Za-z0-9+/=\s]+$/i.test(src) &&
@@ -112,6 +112,34 @@ function isSupportedImageSrc(src: string, source: RepoIconImageSource): boolean
   return url.hostname === 'www.google.com' && url.pathname === '/s2/favicons'
 }
 
+type ImageSrcVerdict = { src: unknown; source: unknown; supported: boolean }
+
+/**
+ * Why: `getRepos()` re-hydrates every repo on every call, and validating one inline data URI means
+ * scanning a 400 KB string twice with a regex and base64-decoding its header. `hydrateRepo` is
+ * handed the *same* persisted `repoIcon` object every time, so the verdict is cached on that object
+ * and dies with it — no cap, no eviction, and nothing retained once a repo or an icon is replaced.
+ *
+ * `src`/`source` are re-checked on a hit, so mutating the persisted icon in place cannot serve a
+ * stale verdict. Both are the identical string references in the steady state, so the compare is a
+ * pointer check, not a 400 KB scan.
+ */
+const imageSrcVerdicts = new WeakMap<object, ImageSrcVerdict>()
+
+function isSupportedImageSrc(
+  candidate: Record<string, unknown>,
+  src: string,
+  source: RepoIconImageSource
+): boolean {
+  const cached = imageSrcVerdicts.get(candidate)
+  if (cached && cached.src === candidate.src && cached.source === candidate.source) {
+    return cached.supported
+  }
+  const supported = computeIsSupportedImageSrc(src, source)
+  imageSrcVerdicts.set(candidate, { src: candidate.src, source: candidate.source, supported })
+  return supported
+}
+
 export function sanitizeRepoIcon(value: unknown): RepoIcon | null | undefined {
   if (value === undefined) {
     return undefined
@@ -146,7 +174,7 @@ export function sanitizeRepoIcon(value: unknown): RepoIcon | null | undefined {
     if (!isRepoIconImageSource(source) || src.length > MAX_REPO_ICON_DATA_URL_LENGTH) {
       return undefined
     }
-    if (!isSupportedImageSrc(src, source)) {
+    if (!isSupportedImageSrc(candidate, src, source)) {
       return undefined
     }
     const label = typeof candidate.label === 'string' ? candidate.label.trim().slice(0, 80) : ''