MaxView

• Rewrite timing.py: snapshot section unchanged; add per-hardware trend tables (rows=tests, cols=last 10 main runs newest-first) with SVG sparklines and heat-map cell backgrounds (green=fastest, red=slowest)
• Wire _build_timing_page to receive all_main_runs_by_hw and metrics_cache so trend data is cached and reused across builds
• Add Timing nav link to all pages (html_primitives.py)
• Force rebuild: all 84 pages updated with new nav + timing trends

• New timing.html: table of test × hardware with wall-clock duration from first/last log timestamps, plus a speedup column (v6e-8/v6e-32 ratio)
• parsers.py: _parse_log_duration() reads log timestamps for true elapsed time
• pages/timing.py: new page builder
• orchestrator.py: wired into global page build
• html_primitives.py: "Timing" nav link added to all pages

Old runs that still have results for 10_shardy_true,

11_optimizer_offload_false, or 13_scan_layers_true now show a amber

DEPRECATED badge next to the test name with a tooltip explaining why

the test was removed. The badge appears in all three comparison tables:

per-mode results, branch-vs-main, and NNX-vs-Linen.

Remove 3 test cases per mode (6 total across Linen + NNX) that are

identical to existing tests because they only explicitly set a flag to

its base.yml default value:

• 10_shardy_true → same as 01_base (shardy=True is default)
• 11_optimizer_offload_false → same as 02_synthetic (False is default)
• 13_scan_layers_true → same as 02_synthetic (True is default)

The non-default counterparts (shardy_false, offload_true, scan_false)

are kept as they test distinct code paths. Total suite: 56 → 50 tests.

Previously every branch run was compared against the current latest main,

so comparisons would silently change whenever a new main landed. Now each

branch run is compared against the most recent main whose timestamp is

at-or-before the branch run's own timestamp, making each comparison a

stable snapshot.

Previously _mvData was loaded once and never re-fetched, so running a

rebuild while the page was open left search with the old index. Now

mvSearchOpen() always kicks off a fetch; the HTTP cache handles the

common case where nothing changed (304), and a fresh response replaces

stale in-memory data after a rebuild.

_html() did a blind str.replace("</table>", "</table></div>") that

appended a stray </div> after every table close, including tables with

attributes (style="width:auto", class="panel-table") used in the tab

panels. Those tables never got the matching opening <div class="table-wrap">

so each stray </div> prematurely closed the enclosing panel div, pushing

the subsequent p/table elements outside the tab.

Fix: replace the two-pass str.replace with a single regex that atomically

wraps only bare <table>...</table> pairs (no attributes), leaving all

attributed tables untouched.

Replaces four separate collapsible sections (System Info, Memory Breakdown,

HBM Utilization, Per-Step Metrics) with a single card containing four tabs:

Summary, System, HBM, Steps. One card, one click to switch — cleaner layout.

Adds four collapsible sections below the summary diff table on every

log-compare page, matching what the individual log summary panel shows:

• System Info: Jaxlib, Hardware, Devices side-by-side
• Memory Breakdown: Total/Output/Temp/Argument for both runs
• HBM Utilization: per-device SVG bar charts + usage tables, one column per run
• Per-Step Metrics: TFLOP/s sparklines side-by-side, then a step-aligned comparison table with Loss Δ and TFLOP/s Δ (green = improvement, red = regression)

All sections are collapsible via <details> with the same smooth CSS animation

already used elsewhere on the site.

New features:

• Inline search overlay: clicking "Search" (or pressing "/") expands a modal with instant client-side search across tests and runs; no separate page navigation required
• Per-test trend pages with TFLOP/s sparklines, flaky test detection, and bisect helper linking last-pass → first-fail commits on GitHub
• Metrics cache (metrics_cache.json) extracts TFLOP/s, tokens/s, loss from run logs for trend visualisation (3600+ entries)
• Global search page + search.json data file
• Shields.io badge endpoint (docs/badge/main.json)
• PR badge on index cards (branch → open PR via gh pr list)
• Sticky filter bar with hardware and status filters on index
• Collapsible branch groups with localStorage persistence
• Branch-level pass-rate sparklines on index
• Summary stats banner (total runs, branches, pass rate)
• Stale data banner when newest run is >3 days old
• Incomplete run badge with hover tooltip
• HBM utilisation bar chart and TFLOP/s sparkline on log pages
• ← Back to run button and History column on run/log pages
• Last-built timestamp footer on every page
• Incremental build system (build_manifest.json; skips unchanged runs)
• watch_site.py: poll-based auto-rebuild on file changes
• check_regressions.py: PASS→FAIL regression alerting with --gh-issue flag
• --max-age-days and --max-runs-per-branch CLI flags on build_site.py

Bug fixes:

• Fixed multiple "All Runs" headers caused by Python implicit f-string concatenation swallowing the join separator
• Fixed run cards split into two chips by replacing nested <a> PR badge with <span onclick=window.open(...)>
• Fixed JS syntax error in search overlay caused by unescaped single quotes in onmouseover handlers; moved hover to CSS instead

fix: and refactor: prefixes are now only shown when scoped to (maxview).

Plain fix:/refactor: commits (infra, scripts) no longer appear on the

Updates page. feat: still included since feature commits tend to be

MaxView-relevant.

• Fix Home/Test Reference/Updates links on log-compare pages (path depth was off by one, resulting in 404 on GitHub Pages).
• Make Metrics section collapsible via <details>; chevron indicator.
• Add prev/next buttons to navigate test cases without returning to the run index.

Store dates as UTC ISO 8601 in meta.json (run_tests_xpk.sh).

Wrap date strings in <time data-utc> elements; a small inline JS

snippet converts them to the browser's local timezone on page load.

• Wrap Updates + Test Reference buttons in .nav-btns flex div so they don't stack on top of each other at position:absolute right:16px
• Add SVG favicon (bar-chart / MV motif) as inline data URI

site: rebuild with nav fix, favicon, and Updates page link

• Font stack: system-ui monospace stack, ligatures, tabular numbers, font smoothing, optimized line-height across body/pre/headings
• Breadcrumb: macOS-style path bar in dedicated .bc-bar strip below header; chip items with hover, › separator, bg matches header (--bg2)
• Header: Test Reference button vertically centered via absolute pos
• Colorize SKIP/SKIPPED results with nan class
• Error/warning log highlighting: Traceback, *Error:, E####, WARNING
• Ghost-run guard in discovery: skip runs with no linen and no nnx data
• Lift _commit_chip and _esc_highlight to module level (dedup)
• Move inline imports (json, re) to module level in parsers/index
• Fix test_reference.py and orchestrator.py to read test plans from runs_dir (not runs_dir.parent)
• Add mkdir(parents=True) guard in _build_log_page before write
• Fix TestParseResultsMd: mock meta.json sibling read so all 57 tests pass

_pick_baseline was falling back to the most-recent baseline of any hardware

when no hw-matched baseline existed. This caused main_20260417_125726 (v6e-32)

to compare against the v6e-8 prev baseline (no v6e-32 prev exists yet).

Fix: return None when no hw-matched baseline — never substitute a different

machine's baseline. Same fix applied to _hw_matched_main on the index page.

All / v6e-8 / v6e-32 toggle buttons appear above the text search bar.

Clicking a hardware button hides cards whose data-hw doesn't match and

collapses branch-groups that become fully hidden. Both filters compose.

• Current main baseline now shows a vs-prev-baseline comparison section
• Commit hashes displayed inline in the section heading for both sides
• Older main baselines show no comparison (avoids confusing main-vs-main)
• Feature branches unchanged: still compare against current main baseline
• Gracefully handles case where no previous baseline exists

Extracted 2046-line monolith into a proper Python package:

maxview/

config.py — site constants

assets.py — CSS, JS assets, test_tooltip_js()

html_primitives.py — _html, _badge, _sanitize_md, etc.

parsers.py — results.md, test plan, log file parsers

discovery.py — run discovery

xplane.py — XPlane profile cache

pages/

index.py — index page

run.py — per-run page

results.py — linen/nnx results page

log.py — individual log page

log_compare.py — side-by-side log compare (unified sync JS)

test_reference.py — test reference page

orchestrator.py — build() / main()

build_site.py becomes a thin shim re-exporting all names for

backward compatibility with tests. All 57 unit tests pass.