A public site.
Two source
trajectories.
The original SpaceX.com source and build system are not publicly available. This experiment reconstructs the observable public site as two Nift projects. It does not claim to migrate or benchmark SpaceX’s private codebase.
Frozen routes and browser evidence first. Publication telemetry second. Independent research; no SpaceX endorsement.
Preserve the visible flight.
A visual, media-heavy marketing site is a different workload from a documentation corpus. Full-bleed video, changing header states, forms, galleries and WebGL trackers all belong to the reconstruction.
- Homepage / settled geometrySeven heading boxes, footer and total body height match both models at 1440×1000, 768×1000, 390×844 and 320×844.
- Tracker / native browser runtimeVehicle and follow-Dragon heading, canvas, telemetry and footer geometry match all four fixtures. Representative pointer drag and keyboard activation work.
- Navigation / scrolling contractMobile down-scroll hides the header; up-scroll restores it. Delivered dropdown positioning and nav-owned menu scrolling are retained.
- Corpus / publication audit829 pages and redirects plus 89 HLS manifests. All 805 reference routes represented; zero unclassified missing local references.
- Forms / local behaviorSource-derived required fields, distinct email grammars, conditional fields and the lead form’s 449/450/500 counter boundaries.
- Page families / representative states712 missions, 143 galleries / 411 slides, 21 crew missions, 22 multi-webcast lists and 3,059 career records are inventoried.
Two ways to own the reconstruction.
Keep repeatable content authoritative.
Structured JSON, prose and native HTML partials maintain mission, homepage and careers content. Shared Nift templates compose them. This is not a Markdown conversion benchmark.
Maintained structured content + partials → Nift dependency evaluation / composition → HTML and HLS publication
Maintain the recovered HTML directly.
Agents edit maintained rendered markup with shared native partials. No recovery script or renderer runs during routine publication.
Maintained rendered HTML + partials → Nift dependency evaluation / composition → HTML and HLS publication
Both retain ordinary browser JavaScript for animation, validation, native/HLS video and WebGL telemetry. Internal links, images, videos, fonts and runtime assets use @path; Nift handles publication paths, not browser physics.
The public reference is not a timing participant.
No private SpaceX build measurements exist in this experiment. The two columns below compare the reconstructed Nift projects only.
Five serialized samples per scenario on an Intel i7-12700H, 20 logical CPUs, Nift 4.10.0 with the maintained automatic thread setting. Medians; zero-based linear scales. Maximum individual process RSS is measured by GNU time, not aggregate memory. No sampled tree-RSS cohort was captured.
Warm full uses nift build --all. Fresh application state clears generated Nift dependency metadata before nift build; it does not flush OS caches. Unchanged uses the same normal production command. Full and fresh build all 918 outputs.
Measure the next edit, including memory.
One mission rebuilds one output; three missions rebuild three; homepage changes rebuild three including aliases. Shared layout, navigation, CSS and JS/cache-key changes rebuild 783 pages. Runtime JSON is already a maintained publication asset and requires no HTML rebuild. A route add or rename builds one output; deletion removes the old output.
These are controlled maintained-source probes. Shared markup/JS probes add harmless comments, CSS adds an unused rule, and metadata adds valid JSON whitespace. Route tracking and source mutation occur before the timed publication command. Every incremental workload reports RSS from those same five samples.
All medians and observed ranges / seconds and MiB
| Scenario | Authored time / s | Rendered time / s | Authored RSS / MiB | Rendered RSS / MiB |
|---|---|---|---|---|
| Warm full | 1.381 (1.337–1.529) | 0.988 (0.840–1.065) | 102.6 (102.1–104.6) | 105.6 (105.1–111.0) |
| Fresh state | 1.355 (1.281–1.369) | 0.844 (0.829–0.875) | 104.4 (102.5–106.0) | 105.1 (101.1–111.3) |
| Unchanged | 0.164 (0.155–0.171) | 0.146 (0.140–0.147) | 7.3 (7.2–7.6) | 7.1 (6.9–7.2) |
| One mission | 0.193 (0.189–0.203) | 0.156 (0.150–0.169) | 9.7 (9.4–10.0) | 8.7 (8.4–9.1) |
| Three missions | 0.243 (0.231–0.279) | 0.182 (0.171–0.200) | 12.1 (11.9–12.2) | 10.7 (10.5–10.9) |
| Homepage | 0.189 (0.175–0.204) | 0.153 (0.149–0.183) | 9.7 (9.1–9.9) | 9.4 (9.0–9.9) |
| Shared layout | 1.470 (1.410–1.539) | 0.924 (0.861–0.953) | 83.4 (82.2–84.4) | 83.5 (83.3–84.1) |
| Navigation | 1.441 (1.386–1.499) | 0.880 (0.873–0.920) | 82.9 (82.1–84.7) | 82.5 (81.9–83.8) |
| Shared CSS | 1.542 (1.535–1.656) | 0.946 (0.913–0.998) | 82.4 (81.6–83.5) | 83.8 (82.3–84.9) |
| JS + cache key | 1.623 (1.605–1.745) | 0.894 (0.877–0.962) | 81.8 (81.3–83.4) | 83.6 (79.5–84.0) |
| Runtime JSON | 0.213 (0.211–0.246) | 0.141 (0.139–0.154) | 7.2 (6.7–7.6) | 7.1 (7.0–7.5) |
| Route added | 0.216 (0.208–0.223) | 0.182 (0.165–0.191) | 9.1 (8.6–9.3) | 8.7 (8.3–8.8) |
| Route renamed | 0.221 (0.198–0.246) | 0.171 (0.154–0.177) | 8.7 (8.6–9.2) | 8.7 (8.3–9.0) |
| Route deleted | 0.209 (0.184–0.221) | 0.167 (0.154–0.173) | 7.3 (6.6–7.8) | 6.9 (6.8–7.2) |
Ten lifecycle correctness checks per project match incremental and forced bytes, including source restoration and absence of renamed/deleted outputs. Fresh GitHub checkouts after pinned setup reproduce all 918 generated files exactly and serve homepage/JS through the normal preview script.
Keep the remaining cost visible.
The rendered model is faster in full publication; it does not uniformly reduce peak RSS. Shared-shell fan-out remains the major iteration cost. Output contains approximately 102 MB HTML, including roughly 81 MB repeated inline CSS, plus 2.18 MB external JavaScript. Inline CSS is already included in HTML bytes, not an additional asset total. Externalizing appropriate shared styles is future work requiring its own parity pass.
No Nift core changes, benchmark-specific cache, thread tuning or speculative optimization layer were introduced. The approximately one-second publication profile did not justify destabilizing the accepted rendering architecture.
Playback reference, not a video archive.
What the observed site needs.
HTML, delivered bundles, images, fonts, hero/background video, posters, thumbnails, tracker textures and captured initial HLS playback windows. Seven pinned maintained-asset releases represent 3,072 paths / approximately 3.858 GB of accepted bytes.
What the download catalogue needs.
The manifest records 61 remote-referenced media entries (about 346.6 GB in declared content lengths). The 320.6+ GiB download-only catalogue remains at original public URLs with route association, role, content type, content length, status and available response metadata. Large payloads were not mirrored just to demonstrate working download links.
Uncaptured HLS tails retain canonical URLs; complete offline streams are not claimed. Original HEVC bytes remain intact, including a clip unsupported by this test browser. External YouTube URLs remain intact; offline parity preview blocks third-party playback. Background videos were not replaced with posters to manufacture parity.
The maintained projects each acquire their required publication assets once. Heavy canonical reference, release transport and benchmark logs stay outside Labs. Normal build/serve does not read sibling reference directories.
Prefer the authored source of truth.
Preferred: Nift spacex — authored source.
Structured authoritative records and shared native templates are easier to review and maintain than coordinating repeated rendered markup.
Preferred: Nift spacex — authored source.
Ordinary iteration is already sub-second. The rendered model’s full-build advantage does not outweigh the clearer content model for this corpus.
If reconstruction effort is already paid, the same preference holds. Nift spacex-agent remains a viable direct-HTML maintenance model and the faster full-publication path. This judgment compares these two reconstructed repositories, never SpaceX’s unseen private source.
Is this kind of site a good Nift fit?
Yes, within the demonstrated public surface: Nift composes a substantial static route corpus cheaply while native browser runtimes retain the media, animation and tracker behavior.
The work also shows the real reconstruction cost, extensive parity work, CSS duplication and unavailable private-service boundaries. It supports Nift’s architectural suitability for this class of public site, not a claim that SpaceX should migrate its private implementation.
Migration-init dogfooding
nift init --migration guidance should explicitly support frozen observable-output reconstruction when source is unavailable, remote-only download media, authored/rendered recovery choices, scoped browser parity and absent incumbent build benchmarks. Those are recorded guidance lessons; no core change was made during this experiment.
Reproduce the flight test.
Source closeouts: authored a467f6f, rendered 1512408. Timing/parity build revisions are separately recorded as 5c94f5e / 0691f43; closeout adds diagnostic scripts, evidence and documentation without changing publication inputs.
Nift spacex — authored repository · Nift spacex-agent — rendered repository
- Frozen public-route inventory ↗
- Public reference and acquisition provenance ↗
- Offline / remote media policy ↗
- Media preservation record ↗
- Pinned maintained asset manifest ↗
- Accepted parity scope and qualifications ↗
- Four-viewport geometry evidence ↗
- Header, dropdown and menu evidence ↗
- All time / RSS samples and distributions ↗
- Authored incremental / forced equality ↗
- Rendered incremental / forced equality ↗
- Authored fresh-clone proof ↗
- Rendered fresh-clone proof ↗
- Publication bytes and repeated CSS ↗
- Authored evaluation and reproduction ↗
- Rendered evaluation and reproduction ↗
- Migration-init lessons ↗
- Complete acquired-reference hash manifest ↗
- Raw time/RSS reports and build logs ↗
- Canonical campaign scope and reference identity ↗
python3 scripts/setup-assets.py # one-time pinned acquisition nift build python3 scripts/serve-preview.py --port 4173 # optional offline-service preview: --offline-services
Final experiment accepted within recorded scope. This report preserves limitations and remaining profiling opportunities; they belong to future tooling work rather than an open-ended polishing campaign.