SHARD MAP DRAFT — the hop-era fleet shard map, computed from JOB-007

DRAFT — pending coordinator review and hop-recon factor calibration. Computed 2026-08-22 from the JOB-007 per-row wall times (docs/phase4/DATA-sweep-row-walltimes.md, windows · corpus 18770d083 · i9-13900K, 2026-08-23), the H10 construction in docs/PLAN-hop-campaign.md §4.3, and the canonical fleet roster in docs/phase4/LANES.md. Every speed factor below is a PROVISIONAL PLACEHOLDER — LANES.md marks historical cross-machine ratios SUSPECT (“the hop shard map’s speed factors come from FRESH same-workload calibration at campaign recon, never from pre-anchor history”). The assignments are the method’s output at the placeholder factors; the map is re-emitted at recon once s_w and k are measured. Nothing in this file is a repo change; the paste-ready section at the end is the proposed insertion, gated on coordinator review.

1. The dataset — JOB-007, parsed

162 rows · 18,569 verdicts · 7,701 i9-seconds total (128.3 min single-machine serial) — the parsed deltas sum to 7,701 s against the sweep’s own 7,697 s aggregate, matching the DATA file’s recorded self-check (the 4 s spread is the derivation’s stated mtime-delta noise).

Statistic	Value
rows	162
total wall	7,701 i9-s (128.3 min)
median row	10 s
mean row	47.5 s
p75 / p90 / p95	16 s / 71 s / 226 s

Distribution shape — extreme right skew. The bulk of the roster is per-row overhead; the wall lives in a handful of rows:

Bucket	Rows	Wall	Share of total
≤ 10 s	95	831 s	10.8 %
11–30 s	41	633 s	8.2 %
31–60 s	8	357 s	4.6 %
61–120 s	6	549 s	7.1 %
121–300 s	6	1,224 s	15.9 %
> 300 s	6	4,107 s	53.3 %

Top-10 heaviest rows — 5,017 s = 65.1 % of the total wall:

Row	Verdicts	Wall
`crypto/dsa`	4	1,317 s
`hash/maphash`	22	898 s
`crypto/tls`	400	659 s
`index/suffixarray`	12	573 s
`archive/zip`	100	354 s
`go/internal/gcimporter`	583	306 s
`go/parser`	173	259 s
`crypto/internal/mlkem768`	12	228 s
`regexp`	45	226 s
`time`	159	197 s

The DATA file’s own header names the two Go-vs-C# inversions this table confirms: crypto/dsa 1,317 s for 4 verdicts and hash/maphash 898 s for 22 — verdict counts are a bad proxy for wall time in both directions (OQ-H5’s exact argument: go/doc/comment carries 10,059 verdicts at 18 s in this reading, archive/zip 100 verdicts at 354 s).

2. Method and inputs

Construction: PLAN-hop-campaign.md §4.3, followed literally —

rows := the 162 JOB-007 rows (re-read at the hop branch tip before dispatch, never carried);
R := the reserved set ∩ rows — all 7 present (index/suffixarray, hash/maphash, crypto/dsa, archive/zip, crypto/tls, go/doc/comment, go/types), 3,956 i9-s (65.9 min), pinned to the i9;
B := the remaining 155 rows, 3,745 i9-s, sorted DESC by t_r (deterministic name tie-break), assigned LPT-greedy: each row to the bin with the smallest current projected local time (load / s_w), the plan’s literal rule;
any bin whose load / s_w exceeds C = 5,400 s (~90 min) splits into sequential shards — at these factors and k = 1, no bin splits at any W;
checksum: rows = R + B = 7 + 155 = 162 — holds for every W below.

Interpretation note for coordinator review: §4.3’s step 3 (recon: P dealt round-robin ASC) and step 4 (bulk: B LPT-greedy DESC) both range over rows \ R. This draft reads the recon deal as an execution-ordering directive, not a second allocation: all 155 non-reserved rows are allocated by LPT, and each worker runs its bin cheapest-first, so the campaign’s opening rows are the cheap ones and k is measured on the first ten per the plan’s own placeholder note. If the coordinator intended recon as a separate dealt tranche, the bins move by single-digit seconds — the cheap rows are 6–10 s each — and the makespans below are unchanged to the minute.

Fleet and provisional speed factors (s_w, i9 = 1.00 — PLACEHOLDERS, all four, to be replaced by the LANES.md calibration pair at hop recon; the 0.35/0.45 values are rough readings-of-record — CLAUDE.md’s “roughly 3–4× the i9” for the i7-5820K class — not measurements of these boxes on this workload):

W	Worker	Silicon	`s_w` (provisional)
3+	i9 (sweeper)	i9-13900K, 16C/24T	1.00 (definition)
3+	R (R-LAPTOP)	Ryzen 7 PRO 6850U, 8C/16T	0.45
3+	coordinator	i7-5820K, 6C/12T	0.35
4+	G (G-LAPTOP)	Ryzen 5 PRO 6650U, 6C/12T	0.35
5	X (fifth engaged machine)	unknown — placeholder silicon	0.35 (assumed laptop-class)

t_r is the JOB-007 wall time × k; k = 1 is assumed throughout this draft (see §4’s caveat — the DATA header describes the measured wall as convert + build + both hosts + compare, but the plan carries k as a recon-measured placeholder and this draft does not preempt that).

3. The assignments

Reserved rows are marked †; the number after each row is its JOB-007 i9-seconds. Load is in i9-seconds; “local wall” divides by the provisional s_w.

W = 3 – makespan 4,289 s local (~71 min)

Worker	s_w	Rows	Load (i9-s)	Local wall	Shards @ ~90 min
i9-13900K (sweeper)	1.00	45	4,272	4,272 s (~71 min)	1
R – R-LAPTOP 6850U	0.45	66	1,928	4,284 s (~71 min)	1
i7-5820K (coordinator)	0.35	51	1,501	4,289 s (~71 min)	1

i9-13900K (sweeper) (45 rows):

index/suffixarray† 573 · hash/maphash† 898 · crypto/dsa† 1317 · archive/zip† 354 · crypto/tls† 659 · go/doc/comment† 18 · go/types† 137 · crypto/aes 9 · crypto/hmac 9 · crypto/sha256 9 · database/sql/driver 9 · debug/plan9obj 9 · encoding/hex 9 · go/constant 9 · hash/crc32 9 · internal/coverage/cformat 9 · internal/coverage/slicereader 9 · internal/cpu 9 · internal/diff 9 · internal/testenv 9 · internal/xcoff 9 · io/ioutil 9 · log/slog/internal/benchmarks 9 · testing/iotest 9 · text/scanner 9 · container/list 8 · crypto/des 8 · crypto/md5 8 · encoding/base32 8 · encoding/base64 8 · hash/fnv 8 · image/color 8 · internal/coverage/slicewriter 8 · internal/itoa 8 · maps 8 · math/cmplx 8 · path 8 · plugin 8 · runtime/internal/math 8 · cmp 7 · container/heap 7 · crypto/internal/alias 7 · hash/adler32 7 · internal/saferio 7 · math/bits 7

R – R-LAPTOP 6850U (66 rows):

go/internal/gcimporter 306 · crypto/internal/mlkem768 228 · time 197 · crypto/rsa 119 · compress/flate 106 · sync/atomic 82 · crypto/internal/edwards25519/field 66 · strings 56 · database/sql 47 · os/exec 43 · crypto/ecdsa 32 · encoding/json 28 · text/template 25 · compress/gzip 21 · regexp/syntax 20 · bytes 18 · debug/buildinfo 18 · sync 18 · crypto/subtle 16 · strconv 16 · crypto/ecdh 15 · crypto/ed25519 15 · image/png 14 · net/http/fcgi 14 · syscall 14 · crypto/rand 13 · expvar 13 · crypto/internal/bigmod 12 · flag 12 · os/signal 12 · debug/gosym 11 · go/ast 11 · go/printer 11 · runtime/debug 11 · debug/macho 10 · encoding/asn1 10 · errors 10 · go/format 10 · hash 10 · internal/abi 10 · internal/fuzz 10 · internal/reflectlite 10 · io/fs 10 · mime/quotedprintable 10 · net/mail 10 · net/url 10 · testing/slogtest 10 · bufio 9 · compress/bzip2 9 · crypto/rc4 9 · encoding/binary 9 · go/token 9 · internal/coverage/pods 9 · internal/dag 9 · io 9 · math 9 · unicode/utf16 9 · crypto/internal/boring 8 · go/version 8 · internal/buildcfg 8 · internal/sysinfo 8 · os/exec/internal/fdtest 8 · text/tabwriter 8 · container/ring 7 · internal/gover 7 · unicode/utf8 6

i7-5820K (coordinator) (51 rows):

go/parser 259 · regexp 226 · internal/godebugs 177 · encoding/pem 105 · mime/multipart 71 · archive/tar 60 · encoding/xml 54 · math/rand/v2 34 · math/rand 31 · html/template 28 · go/importer 20 · go/internal/srcimporter 19 · compress/zlib 18 · crypto/elliptic 16 · sort 16 · crypto 15 · image/gif 14 · internal/types/errors 14 · net/rpc/jsonrpc 14 · context 13 · path/filepath 13 · debug/elf 12 · image 12 · debug/dwarf 11 · go/internal/gccgoimporter 11 · image/jpeg 11 · crypto/internal/hpke 10 · encoding/csv 10 · fmt 10 · go/scanner 10 · image/draw 10 · internal/profile 10 · internal/zstd 10 · mime 10 · net/textproto 10 · runtime/metrics 10 · text/template/parse 10 · compress/lzw 9 · crypto/sha512 9 · go/build/constraint 9 · internal/coverage/cmerge 9 · internal/singleflight 9 · log 9 · testing/quick 9 · crypto/sha1 8 · hash/crc64 8 · internal/fmtsort 8 · net/http/internal/ascii 8 · unicode 8 · encoding/ascii85 7 · runtime/internal/sys 7

Checksum: 162 rows assigned = 7 reserved + 155 bulk (every roster row named exactly once).

W = 4 – makespan 3,956 s local (~66 min)

Worker	s_w	Rows	Load (i9-s)	Local wall	Shards @ ~90 min
i9-13900K (sweeper)	1.00	7	3,956	3,956 s (~66 min)	1
R – R-LAPTOP 6850U	0.45	61	1,468	3,262 s (~54 min)	1
i7-5820K (coordinator)	0.35	47	1,138	3,251 s (~54 min)	1
G – G-LAPTOP 6650U	0.35	47	1,139	3,254 s (~54 min)	1

i9-13900K (sweeper) (7 rows):

index/suffixarray† 573 · hash/maphash† 898 · crypto/dsa† 1317 · archive/zip† 354 · crypto/tls† 659 · go/doc/comment† 18 · go/types† 137

R – R-LAPTOP 6850U (61 rows):

go/parser 259 · regexp 226 · crypto/rsa 119 · encoding/pem 105 · crypto/internal/edwards25519/field 66 · encoding/xml 54 · os/exec 43 · math/rand 31 · html/template 28 · go/importer 20 · go/internal/srcimporter 19 · debug/buildinfo 18 · crypto/subtle 16 · strconv 16 · crypto/ed25519 15 · image/png 14 · net/rpc/jsonrpc 14 · crypto/rand 13 · path/filepath 13 · debug/elf 12 · os/signal 12 · go/ast 11 · image/jpeg 11 · runtime/debug 11 · encoding/asn1 10 · fmt 10 · hash 10 · internal/abi 10 · internal/profile 10 · io/fs 10 · net/mail 10 · runtime/metrics 10 · text/template/parse 10 · compress/bzip2 9 · crypto/hmac 9 · crypto/sha512 9 · encoding/binary 9 · encoding/hex 9 · go/token 9 · internal/coverage/cmerge 9 · internal/cpu 9 · internal/diff 9 · internal/testenv 9 · io/ioutil 9 · math 9 · text/scanner 9 · unicode/utf16 9 · crypto/internal/boring 8 · encoding/base32 8 · hash/crc64 8 · internal/buildcfg 8 · internal/coverage/slicewriter 8 · internal/sysinfo 8 · net/http/internal/ascii 8 · plugin 8 · text/tabwriter 8 · cmp 7 · crypto/internal/alias 7 · internal/gover 7 · math/bits 7 · unicode/utf8 6

i7-5820K (coordinator) (47 rows):

crypto/internal/mlkem768 228 · time 197 · compress/flate 106 · mime/multipart 71 · strings 56 · math/rand/v2 34 · encoding/json 28 · compress/gzip 21 · bytes 18 · sync 18 · sort 16 · crypto/ecdh 15 · internal/types/errors 14 · syscall 14 · expvar 13 · flag 12 · debug/dwarf 11 · go/internal/gccgoimporter 11 · crypto/internal/hpke 10 · encoding/csv 10 · go/format 10 · image/draw 10 · internal/reflectlite 10 · mime 10 · net/textproto 10 · testing/slogtest 10 · compress/lzw 9 · crypto/rc4 9 · database/sql/driver 9 · go/build/constraint 9 · hash/crc32 9 · internal/coverage/pods 9 · internal/dag 9 · internal/xcoff 9 · log 9 · testing/iotest 9 · container/list 8 · crypto/md5 8 · encoding/base64 8 · hash/fnv 8 · internal/fmtsort 8 · maps 8 · os/exec/internal/fdtest 8 · runtime/internal/math 8 · container/heap 7 · encoding/ascii85 7 · internal/saferio 7

G – G-LAPTOP 6650U (47 rows):

go/internal/gcimporter 306 · internal/godebugs 177 · sync/atomic 82 · archive/tar 60 · database/sql 47 · crypto/ecdsa 32 · text/template 25 · regexp/syntax 20 · compress/zlib 18 · crypto/elliptic 16 · crypto 15 · image/gif 14 · net/http/fcgi 14 · context 13 · crypto/internal/bigmod 12 · image 12 · debug/gosym 11 · go/printer 11 · debug/macho 10 · errors 10 · go/scanner 10 · internal/fuzz 10 · internal/zstd 10 · mime/quotedprintable 10 · net/url 10 · bufio 9 · crypto/aes 9 · crypto/sha256 9 · debug/plan9obj 9 · go/constant 9 · internal/coverage/cformat 9 · internal/coverage/slicereader 9 · internal/singleflight 9 · io 9 · log/slog/internal/benchmarks 9 · testing/quick 9 · crypto/des 8 · crypto/sha1 8 · go/version 8 · image/color 8 · internal/itoa 8 · math/cmplx 8 · path 8 · unicode 8 · container/ring 7 · hash/adler32 7 · runtime/internal/sys 7

Checksum: 162 rows assigned = 7 reserved + 155 bulk (every roster row named exactly once).

W = 5 – makespan 3,956 s local (~66 min)

Worker	s_w	Rows	Load (i9-s)	Local wall	Shards @ ~90 min
i9-13900K (sweeper)	1.00	7	3,956	3,956 s (~66 min)	1
R – R-LAPTOP 6850U	0.45	47	1,124	2,498 s (~42 min)	1
i7-5820K (coordinator)	0.35	36	873	2,494 s (~42 min)	1
G – G-LAPTOP 6650U	0.35	36	874	2,497 s (~42 min)	1
X – fifth engaged machine	0.35	36	874	2,497 s (~42 min)	1

i9-13900K (sweeper) (7 rows):

index/suffixarray† 573 · hash/maphash† 898 · crypto/dsa† 1317 · archive/zip† 354 · crypto/tls† 659 · go/doc/comment† 18 · go/types† 137

R – R-LAPTOP 6850U (47 rows):

go/parser 259 · time 197 · sync/atomic 82 · archive/tar 60 · encoding/xml 54 · math/rand/v2 34 · html/template 28 · regexp/syntax 20 · compress/zlib 18 · crypto/elliptic 16 · strconv 16 · image/gif 14 · net/http/fcgi 14 · context 13 · path/filepath 13 · image 12 · go/ast 11 · image/jpeg 11 · crypto/internal/hpke 10 · errors 10 · hash 10 · internal/fuzz 10 · internal/reflectlite 10 · mime/quotedprintable 10 · runtime/metrics 10 · compress/bzip2 9 · compress/lzw 9 · crypto/sha256 9 · encoding/binary 9 · go/token 9 · internal/coverage/cmerge 9 · internal/coverage/slicereader 9 · internal/singleflight 9 · io/ioutil 9 · testing/iotest 9 · testing/quick 9 · crypto/des 8 · encoding/base32 8 · hash/fnv 8 · internal/fmtsort 8 · internal/itoa 8 · net/http/internal/ascii 8 · runtime/internal/math 8 · container/heap 7 · container/ring 7 · internal/gover 7 · unicode/utf8 6

i7-5820K (coordinator) (36 rows):

regexp 226 · internal/godebugs 177 · mime/multipart 71 · database/sql 47 · encoding/json 28 · compress/gzip 21 · bytes 18 · crypto/subtle 16 · crypto/ecdh 15 · internal/types/errors 14 · crypto/rand 13 · debug/elf 12 · debug/dwarf 11 · go/printer 11 · encoding/asn1 10 · go/format 10 · internal/abi 10 · io/fs 10 · net/textproto 10 · text/template/parse 10 · crypto/hmac 9 · database/sql/driver 9 · go/build/constraint 9 · internal/coverage/cformat 9 · internal/dag 9 · internal/xcoff 9 · log/slog/internal/benchmarks 9 · unicode/utf16 9 · crypto/sha1 8 · hash/crc64 8 · internal/coverage/slicewriter 8 · math/cmplx 8 · plugin 8 · cmp 7 · crypto/internal/alias 7 · internal/saferio 7

G – G-LAPTOP 6650U (36 rows):

go/internal/gcimporter 306 · compress/flate 106 · crypto/internal/edwards25519/field 66 · os/exec 43 · math/rand 31 · go/importer 20 · debug/buildinfo 18 · sort 16 · crypto/ed25519 15 · net/rpc/jsonrpc 14 · expvar 13 · flag 12 · debug/gosym 11 · runtime/debug 11 · encoding/csv 10 · go/scanner 10 · internal/profile 10 · mime 10 · net/url 10 · bufio 9 · crypto/rc4 9 · debug/plan9obj 9 · go/constant 9 · internal/coverage/pods 9 · internal/diff 9 · io 9 · math 9 · container/list 8 · crypto/internal/boring 8 · encoding/base64 8 · image/color 8 · internal/sysinfo 8 · os/exec/internal/fdtest 8 · text/tabwriter 8 · encoding/ascii85 7 · math/bits 7

X – fifth engaged machine (36 rows):

crypto/internal/mlkem768 228 · crypto/rsa 119 · encoding/pem 105 · strings 56 · crypto/ecdsa 32 · text/template 25 · go/internal/srcimporter 19 · sync 18 · crypto 15 · image/png 14 · syscall 14 · crypto/internal/bigmod 12 · os/signal 12 · go/internal/gccgoimporter 11 · debug/macho 10 · fmt 10 · image/draw 10 · internal/zstd 10 · net/mail 10 · testing/slogtest 10 · crypto/aes 9 · crypto/sha512 9 · encoding/hex 9 · hash/crc32 9 · internal/cpu 9 · internal/testenv 9 · log 9 · text/scanner 9 · crypto/md5 8 · go/version 8 · internal/buildcfg 8 · maps 8 · path 8 · unicode 8 · hash/adler32 7 · runtime/internal/sys 7

Checksum: 162 rows assigned = 7 reserved + 155 bulk (every roster row named exactly once).

4. Analysis

4.1 What dominates

Six rows over 300 s carry 53.3 % of the whole roster’s wall; the top ten carry 65.1 %. Five of those six are already in the plan’s reserved set (crypto/dsa 1,317, hash/maphash 898, crypto/tls 659, index/suffixarray 573, archive/zip 354); the sixth, go/internal/gcimporter (306 s), is the largest bulk row and lands first in every LPT deal. The reserved pin is quantitatively right: crypto/dsa alone on a 0.35-factor box is 3,763 s (62.7 min) — nearly the entire W=4 makespan in one row. No slow worker can carry any of the top four rows without becoming the campaign’s critical path.

Two reserved rows earn their pin for reasons the wall-time column does NOT show, and the map must not “optimize” them out at recon: go/doc/comment reads 18 s here but carries 10,059 verdicts and spawns go build throughout TestStd — its cost under -test-action all (the k regime) will diverge hardest from this table; and go/types read 137 s here against 364 s in the plan’s own reserved table — the row moves between readings, which is itself the argument for pinning it to the fastest, best-characterized box.

4.2 The structural finding: at W ≥ 4 the makespan IS the reserved set

W	Makespan	Binding bin	Bulk bins finish at
3	4,289 s (~71.5 min)	all three balanced (4,272–4,289 s)	—
4	3,956 s (~65.9 min)	the i9’s reserved set, alone	~3,250 s (~54 min)
5	3,956 s (~65.9 min)	the i9’s reserved set, alone	~2,497 s (~42 min)

At W=3 the fleet is capacity-bound: makespan ≈ total / Σs_w = 7,701 / 1.80 = 4,278 s, and LPT lands within 11 s of that bound (the i9 takes 38 cheap bulk rows on top of its reserves). From W=4 the picture inverts: the three non-i9 workers absorb the entire 3,745 s bulk and still finish ~18 % below the i9’s floor, LPT gives the i9 zero bulk rows, and the campaign’s floor is the reserved set’s 3,956 s run serially on the i9. Adding the fifth worker buys no makespan at all — it only drops the bulk bins from ~54 to ~42 min. W=5’s value is margin and re-deal resilience (a reboot costs a smaller shard), not speed.

Two levers if ~66 min needs to shrink, both coordinator decisions, neither assumed here: (1) the i9’s declared capacity is 3 concurrent worktrees — an LPT packing of the 7 reserved rows into 3 parallel worktrees reaches ~1,369 s (≈ 23 min) on the reserved leg, at the cost of running reserved rows under concurrent load, which is exactly the condition their $longTimeouts floors exist to survive (crypto/dsa measured 2,444 s under concurrent gates); (2) peel go/types + go/doc/comment (the two cheap reserves) to a bulk bin, saving 155 s of floor — barely worth the exception to “never sharded blind”.

4.3 Is the ~90-minute shard target achievable?

Yes at every W — with k = 1 and these factors, no bin splits. The margins, as k-thresholds (the multiplier at which a bin first exceeds C = 5,400 s and must split):

W	Largest local bin	Headroom vs 90 min	Splits when k ≥
3	4,289 s	1,111 s	1.26
4	3,956 s (i9 reserved)	1,444 s	1.37
5	3,956 s (i9 reserved)	1,444 s	1.37

So the target survives a convert-and-build multiplier of up to ~1.26 at W=3 and ~1.37 at W≥4 before any bin becomes two sequential shards. If recon measures k materially above that, the construction’s step 5 splits the affected bins mechanically — the map does not need redesigning, only re-emitting. (Whether k ≈ 1 is genuinely open: the DATA header says the measured wall already covers convert + build + both test hosts + compare, which argues k ≈ 1, but the plan prices k as a recon measurement on the first ten rows and this draft defers to that.)

4.4 Sensitivity to the suspect speed factors

Perturbing the provisional factors at W=4 (LPT re-run per scenario):

Scenario	Makespan	Δ vs base
base (i9 = 1.00, R = 0.45, i7 = 0.35, G = 0.35)	3,956 s (65.9 min)	—
slow laptops (R = 0.35, G = 0.25)	3,956 s	+0 %
slow coordinator (i7 = 0.25)	3,956 s	+0 %
fast laptops (R = 0.55, G = 0.45)	3,956 s	+0 %
everything slow (R = 0.35, i7 = 0.25, G = 0.25)	4,176 s	+6 %
i9 degraded 20 % (i9 = 0.80)	4,945 s (82.4 min)	+25 %

(The table’s scenarios re-run LPT at the true factors — i.e. a map built knowing them. The “slow laptops” row lands at +0 % by a hair: the three bulk workers’ combined capacity at the i9’s 3,956 s floor is 0.95 × 3,956 = 3,758 i9-s against the 3,745 i9-s bulk load — 99.7 % utilization.)

The headline: at W ≥ 4 the makespan is only mildly sensitive to the suspect factors — the bulk bins finish ~18 % below the i9’s reserved floor, so a mis-calibrated laptop factor mis-balances bins that were not the critical path anyway. The costlier case is executing a map built on a wrong factor: a worker sized at 0.45 that truly runs at 0.35 finishes its bin ~29 % later than projected (54 → ~70 min — briefly the critical path, lifting the makespan ~6 % over the floor, still well under C). Mis-calibration therefore costs projection accuracy, idle time and single-digit-percent finish slip — never the 90-min target. The one factor that matters is the i9’s own throughput (reboot recovery, concurrent-job load, thermal state): 20 % degradation moves the makespan 25 %, because the i9 is the floor. At W=3 the insensitivity does NOT hold — the makespan there is capacity-bound (≈ total / Σs_w), so every factor error passes straight through; a W=3 dispatch should wait for real calibration numbers, a W≥4 dispatch tolerates the placeholders.

Corollary for OQ-H6’s fallback: if the i9 drops out entirely, the reserved set cannot be re-dealt at these budgets — crypto/dsa + hash/maphash + crypto/tls + index/suffixarray on a 0.45 box is ~7,600 s local (~2.1 h) before the floors are raised. The re-deal rule in §6 lens 3 (raise -TestTimeout, never re-deal at the same budget) is the operative defense; the shard map itself should name R as the reserved-set fallback since it carries the largest non-i9 factor.

5. Ready-to-paste — proposed insertion for `PLAN-hop-campaign.md` §4.3

Insert after the map-construction code block and its symbol table, before “Illustrative widths over 162 rows” (which it supersedes — the illustrative table’s even-split arithmetic is now replaced by a computed deal; the coordinator may retire or keep that table). Marked DRAFT per the header; the full per-row lists land in docs/phase4/SHARDMAP-go1.23.12.md at dispatch time per the construction’s step 6.

### 4.3.1 The computed map at the JOB-007 reading — DRAFT, pending recon calibration

> **DRAFT (2026-08-22) — pending coordinator review and hop-recon factor calibration.** The
> construction above, executed against the anchor's per-row wall times
> ([`phase4/DATA-sweep-row-walltimes.md`](phase4/DATA-sweep-row-walltimes.md), windows · corpus
> `18770d083` · i9-13900K, JOB-007: 162 rows, 18,569 verdicts, **7,701 i9-s total**). ⚠ **Every
> `s_w` below is a provisional placeholder** — LANES.md's roster marks historical cross-machine
> ratios SUSPECT, so the factors here are class-level readings (i9 = 1.00 by definition,
> R 6850U = 0.45, i7-5820K = 0.35, G 6650U = 0.35, any fifth worker = 0.35 assumed), to be
> replaced by the calibration pair's fresh numbers at recon and the map re-emitted. `k` is
> assumed 1 pending its recon measurement. The full per-row deal (every row named exactly once,
> checksum 162 = 7 reserved + 155 bulk) is emitted to `SHARDMAP-go1.23.12.md` at dispatch.

**The reserved set is the makespan from W = 4 up.** R's seven rows sum to **3,956 s (65.9 min)**
serial on the i9 — and at W ≥ 4 the remaining 155 rows (3,745 i9-s) fit on the other workers
~18 % under that floor, so LPT hands the i9 zero bulk rows and the campaign finishes when the
reserved set does:

| `W` | Fleet | Makespan (local wall) | Binding bin | Bulk bins finish |
|:--:|:--|--:|:--|--:|
| 3 | i9 + R + coordinator | **~4,289 s (71.5 min)** | all three balanced | — |
| 4 | + G | **~3,956 s (65.9 min)** | i9's reserved set alone | ~54 min |
| 5 | + one engaged machine | **~3,956 s (65.9 min)** | i9's reserved set alone | ~42 min |

Findings the numbers force, stated before dispatch:

1. **C = 90 min holds at every W at k = 1 — no bin splits.** The split thresholds: a bin first
   exceeds C at k ≥ 1.26 (W=3) / k ≥ 1.37 (W≥4). If recon's k lands above that, step 5 splits
   mechanically; the map re-emits, it does not redesign.
2. **At W ≥ 4 the makespan is only mildly sensitive to the suspect factors** (±0.10 on every
   non-i9 factor moves it 0–6 %) because the bulk bins are not the critical path;
   mis-calibration costs projection accuracy and single-digit finish slip, never the 90-min
   target. **At W = 3 it is capacity-bound and every factor
   error passes through** — a W=3 dispatch waits for real calibration; W≥4 tolerates
   placeholders. The one live sensitivity is the i9 itself: 20 % degradation → +25 % makespan.
3. **W = 5 buys margin, not makespan** (bulk bins 54 → 42 min). Engage a fifth machine for
   re-deal resilience against the i9's ~daily reboots, not for speed.
4. **Six rows carry 53 % of the wall**, five of them already reserved; `crypto/dsa` alone on a
   0.35 box would be 62.7 min. The pin is arithmetic, not caution. Two reserves are pinned for
   reasons this table cannot show: `go/doc/comment` (18 s here, but 10,059 verdicts and spawns
   `go build` throughout `TestStd` — the row most exposed to k) and `go/types` (137 s here vs
   364 s at its prior reading — it moves between readings).
5. **Reserved-set fallback (OQ-H6): R**, the largest non-i9 factor — at raised budgets only
   (~2.1 h local for the top four rows; never re-dealt at the i9's budgets, per §6 lens 3).
6. If ~66 min must shrink: the i9's 3-worktree capacity packs the reserved set to ~1,369 s
   (~23 min) in 3 parallel worktrees — at the cost of running exactly the `$longTimeouts` rows
   under concurrent load (crypto/dsa measured 2,444 s under concurrent gates). A coordinator
   trade, not a default.

Appendix — reproduction

⚠ AMENDMENT (2026-08-24) — this file and shardmap.py are BANKED, and the generator has moved on. “This scratchpad directory” is now docs/phase4/hopA-inputs/ (e0d8930e1, banked as-found). emit_md.py was never banked — it existed only in the session that wrote this file, and it is gone; shardmap.py alone is the reproduction path, and the tables below are its output rather than emit_md.py’s. shardmap.py has since changed in one load-bearing way (549b4e556): it derives its reserved set from run-validated-sweep.ps1’s $longTimeouts at generation time instead of carrying the copied list this draft was computed with — so re-running it today yields a larger reserved set than the 7 rows below, before any factor calibration. Read the deal here as the method’s output at the placeholder factors and the drifted reserved set of its day.

Computed by shardmap.py / emit_md.py in this scratchpad directory (Python 3; parses the fenced JOB-007 table from docs/phase4/DATA-sweep-row-walltimes.md, asserts 162 rows and the 7 + 155 checksum, runs the §4.3 construction literally — reserved pin, DESC LPT to smallest load / s_w, deterministic name tie-breaks — then the sensitivity scenarios of §4.4). The LPT deal is deterministic: two coordinators running it against the same DATA table and factor set emit byte-identical bins.