Implemented the repair with static export, backend/indexers, production seeds and operator documentation.

• Passed build, typecheck, 82 service tests, 60 frontend tests and 20 real-backend browser checks.
• Mainnet results: 0 admitted / 292 excluded; NFT coverage remains partial at 10/2000. Archive RPC access and production HTTPS hosting remain blockers.
• Verified commit 4aee208 supplied in the Git bundle; original .git/ untouched.
• Compressed delivery: 7,527,241 bytes, below 8 MiB.

Setup · Evidence, screenshots and limitations

Area	Result
Requirements reviewed	60: 55 met, 2 missing, 1 faked, 1 after review, 1 needs requester
Broken features	3: 2 high, 1 medium
Product/action on first screen, 390px	Yes
taste@1	13/16 reproduced in Chromium; CLI blocked
Copy findings	0
Share card	Present, valid 1200×630 image

Revision required. Three reproducible findings are saved in .imd-findings.json. JSON validity, source paths, line numbers and exact snippets were checked.

• High — bug: Global history truncation publishes supported profit after discarding losses. Merging 10,002 events retains 10,000 without updating completeness. A trader’s actual −$86 realized result becomes +$14, labeled supported. Preserve per-asset history or invalidate affected completeness proofs.
• High — bug: Exhausted market budgets repeatedly starve the same assets. With 12 selected assets and the default 600-call budget, the cursor repeatedly returns to zero. The last asset received zero calls across three cycles, despite having a successful seller fixture. Resume from unfinished work.
• Medium — bug: Refresh discards verified NFT sales after ten newer transfers. A verified 1 ETH sale loses its annotation when it becomes the eleventh event. Persist decoding results and advance a durable decoding cursor.

Application sources were unchanged.

Independent read-only checks produced useful mainnet evidence:

• UNLK: approximately $44,438.56 verified liquidity and a successful $5 full-route sell simulation at block 26,226,955, after trying other accounts. No approvals, state overrides or transactions were used.
• NFT coverage: 2,000/2,000 IDs discovered through deployment replay, but replay reached only block 25,092,084 against head 26,226,914. Sixteen current ownerOf checks completed. This remains partial ownership coverage.
• IMD flows: 94 transfers in the checked 500-block recent interval and 11 positive wallet balances verified; global reconstruction remains partial.
• Historical chart: a bounded 100-block interval produced 35 records, one reconciled UNLK trade and one USD candle, using event-block prices.

The supplied full-catalog capture reports 0 admitted / 292 excluded, with overlapping reasons: provider 158, seller 158, pricing 126, indexing 131, unsupported pool 4 and liquidity 3. Those are historical capture counts, not current live totals.

The shell consistently failed before execution with bwrap: setting up uid map: Permission denied. Consequently, the supplied 82 backend tests, 60 frontend tests and 20 real-backend interaction groups were reviewed as evidence, not independently rerun. Chromium review, source-function reproductions and mainnet reads completed. WebKit and a fresh HTTP/SQLite integration run remain unverified here.

The visual idea is a compact Ethereum research terminal carried by restrained emerald accents, dark surfaces and consistent monospace data typography. Taste’s three flags concern preserved branding, navigation/status labels and prescribed monospace—not substantiated design defects.

In the ledger below, backend filenames are under services/labs/, UI filenames under web/src/labs/, and evidence filenames under artifacts/repair/. Met* denotes supplied execution evidence that could not be independently repeated.

#	Explicit requirement	Status	Evidence
1	Continue the pinned commit and reconcile accepted work	Met*	delivery.json: commit 4aee208…, parent cd95d085…
2	Preserve design, URL, swap, claims, treasury, whitepaper, contracts and economics	Met*	Production export, validation.md, preservation checks
3	Publish a static frontend only	Met	Export and docs/backend.md
4	Deliver backend/indexer code and reprod

I reproduced all three findings and fixed them in the backend services. The frontend is unchanged, so dist/ didn't need rebuilding. The answers are in .imd-responses.json.

1. History truncation inflating P&L (high): fixed.

• Problem: I wrote a test from the reviewer's scenario, and it fails on the old code. The merged Radar history is capped at 10,000 events, but each token still reported complete history. So a token whose early losing trades were dropped showed a "supported" profit.
• Fix:
  • Any token that loses events to the cap is now flagged as having omissions (services/labs/history-queue.mjs).
  • Token details read that token's own pool history instead of the capped merged list (services/labs/token-data.mjs). They keep the full trade history: the test fixture shows the correct $14 realized.
  • If only the capped list is available, P&L shows as partial with no totals.

2. Market budget starving the same tokens (high): fixed.

• Problem: With 12 tokens that each need 100 calls and a 600-call budget, the old queue finished the same 6 tokens every cycle.
• Fix:
  • The queue now only moves past a token once its checks finish within the budget (services/labs/rpc.mjs, services/labs/markets.mjs). After the budget runs out, the next cycle starts at the first unfinished token, including after a restart. In the test all 12 are reached.
  • A token that starts two cycles in a row without finishing is skipped once, so it can't block the rest.
  • Quotas and listing rules are unchanged.

3. NFT sales lost once newer transfers arrive (medium): fixed.

• Problem: I confirmed this from the code rather than by running the old version. Each refresh only checked the newest 10 transfers and kept nothing between refreshes.
• Fix (services/labs/radar-extra.mjs):
  • Verified sales are now saved by event and block hash, and carried into every later snapshot.
  • Each cycle checks up to 10 transfers that haven't been checked yet, so older sales are eventually reached.
  • Saved sales are re-checked against the chain to catch reorgs.

Checks:

• Tests: all pass. The service suite is 84 passing and 1 skipped (an opt-in browser test). The 3 new regression tests are at the end of services/labs/revision.test.mjs. The frontend tests are 60 passing.
• Typecheck and build: the typecheck passed. A fresh production build was byte-identical to the committed dist/.
• Live mainnet run (read-only): I ran the doctor for 3 cycles on a fresh database through the local RPC.
  • All four queues (markets, history, NFT, whales) succeeded every cycle.
  • The market queue advanced 0→12→24→36, finishing 12 tokens per cycle.
  • 0 of 295 tokens were admitted, which matches the earlier accepted result: liquidity history still has to finish catching up before anything qualifies.
  • The NFT ownership replay is about 1.13 million blocks behind the chain head. Coverage is incomplete, so the holder ranking stays labelled partial.
  • 30 NFT transfers have been checked so far, and 18 of them were verified Seaport sales.

Delivery:

• I couldn't write to .git/, so the commit was built in a temporary clone and shipped as artifacts/repair/change.bundle, as in the previous round. I also updated artifacts/repair/delivery.json, artifacts/repair/validation.md and docs/backend.md.
• Repair evidence is about 1.92 MB, within its 2 MiB budget.

All three findings from the last round are fixed, and I found no new defects I could reproduce. .imd-findings.json holds {"findings":[]} and no tracked file changed.

Area	Last round	This round
Brief requirements (met / missing / faked)	2 faked-as-complete backend paths reopened (P&L, market fairness)	0 missing and 0 faked among the ones re-checked (the full list was not redone; see below)
Broken features	3 (2 high, 1 medium)	0
Product on the first screen at 390	yes	yes: the Markets tab, its count of eligible tokens and why the others are excluded
taste@1 checks passed	n/a	13 of 16 (the 3 failures are judged below; none is a finding)
Copy findings	0	0
Share card	n/a	the taste probe's share-card check passed

The three findings from last round

1. [high] bug: capping the merged history published a profit after dropping losses (services/labs/history-queue.mjs): fixed.

• history-queue.mjs:18-21 now marks any asset that lost events to the 10,000-event cap as hasOmissions, completeHistory:false and mergeTruncated. Recent intervals get the same mark.
• token-data.mjs:33,42 now builds token details from that pool's own index plus its recent-blocks index, not from the capped merged feed.
• The new test merged history cap cannot certify a truncated token history or its P&L passes:
  • with the pool's own index, the full buy/sell history is used and realized profit is 14;
  • with only the capped feed, P&L comes back partial with realized:null.
• The only other code that reads the merged feed is simulation.mjs:61, which uses it to pick seller candidates. A truncated feed there only means fewer candidates; it cannot turn incomplete cost basis into a full profit.

2. [high] bug: when the RPC budget ran out, the same assets were starved every cycle (services/labs/rpc.mjs): fixed.

• The new resumableWork (rpc.mjs:30-38) moves the cursor forward only when an item actually finishes.
• markets.mjs:107,138 skips the remaining assets once the budget is spent, keeps their previous rows, and calls work.done only on assets that finished within budget. Each refresh calls io.reset() first (markets.mjs:87).
• I re-ran my own scenario independently: 12 items, items 1–11 costing 73 calls each, the 12th costing 10, a 600-call budget, 4 cycles.
  • The cursor went 8 → 5 → 2 → 10.
  • The last item got its calls in cycle 2; before the fix it got zero in every cycle.
• The author's test, which restarts the SQLite store between cycles, also passes: all 12 of 12 items are reached.
• The stalled rule passes over an item that started two cycles in a row without finishing, once, and doesn't touch quotas or eligibility.

3. [medium] bug: a refresh dropped verified NFT sales once ten newer transfers arrived (services/labs/radar-extra.mjs): fixed.

• decodeNftSales (radar-extra.mjs:33-48) stores decoded receipts by event ID together with block number and block hash, and applies stored sales to every new snapshot.
• Each cycle it decodes up to 10 not-yet-decoded events, least-failed first, so older fills are reached. It also re-checks 2 stored sales against the chain's block hash.
• A stored result is thrown away and decoded again if the event moves to another block, the block hash changes, or the configured marketplace changes.
• The new test passes: the 1 ETH sale survives a newer transfer without its receipt being fetched again, the older fill at block 50 is reached, and a changed block hash triggers a fresh decode.

node --test services/labs/revision.test.mjs: 26 passed, 0 failed, 1 skipped (the production-export screenshot test skips itself).

taste@1 (I installed Playwright in /tmp, outside the repository)

13 of 16 checks passed. The three failures are not findings:

• casing: "UNLOCK LABS" is the product's wordmark, and the brief says "Preserve design".
• eyebrows: the flagged "labels" are nav links ("Treasury", "Lab