b86bb00a554cbaaae0def67e02fc73d6c90d426c
17
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
  jpmschweitzerandClaude Opus 4.6
|
531bb1df5c |
refactor(assets): atlas naming — dedup, compass filter, river vocab, infra names (#853)
Addresses five of seven subtasks from atlas-generator-refinement-notes:
- Cross-body dedup: corpus keyed by (corridor, feature_type) instead of
(system_id, feature_type), seeded from existing atlas_* rows so re-runs
don't collide with already-committed names. §1, §2.
- Empty-name fallback for mountain ranges when Gemma returns fewer names
than needed ({body_proper} Range {i+1}). §2.
- Suffix monotony detection: flags bodies where >40% of mountain names
share a trailing word (warning only — batch pipeline has no voice
access for auto-fix). §3. Follow-up in #886.
- Compass-direction ban: build_batch_prompt explicitly forbids
"Eastern/Northern/Western X" in few-shot instructions. §4.
- River vocabulary filter: is_valid_name rejects "X Flow" / "X Current"
when feature_type="river" — these are ocean terms bleeding through. §6.
- Infrastructure naming: deterministic post-pass assigns "{CityA}–{CityB}
{corridor_suffix}" to unnamed roads and railroads (Corridor/Road/Estrada/
Strasse/Track by corridor). §7.
Cultural-history prompt threading (§5) remains as existing corridor_substyles
refill mechanism; explicit cultural-history blurb deferred to #886.
naming_core bumped to v0.3.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
jpmschweitzerandClaude Opus 4.6
|
160c3f1853 |
fix(tooling): PR #130 round 2 — stale strings, O(1) dedup, vestigial scope (#833)
- Fix "Gemma 2 GGUF" in user-facing error message (line 2073) - Fix gemma2.gguf in docstring usage example (line 32) - Fix O(N) _is_duplicate: pre-build lowercase shadow sets for O(1) lookup - Expand vestigial note to enumerate full ~750-line dead island boundaries Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
8ba3a1d0a7 |
fix(tooling): address PR #130 review findings (#833)
Hoshe: - Mark --dump-prompts / name_feature() as vestigial with TODO note - Fix --refresh help string: 200 → 1000 (matches actual default) - Fix _RIVER_POOLS comment numbering: Pool 6 before Pool 5 → correct order - Remove dead first-pass code in fix_fewshot_bleed.py - _CAPTURE_FILE leak noted in vestigial TODO Tyre: - Fix stale "Gemma 2" strings in banner, argparse description, model help - Note dead code for cleanup pass (name_feature ~700 lines) Hoshe (prune): - prune_atlas_features.py: named features sort before unnamed, preventing silent discard of hand-authored names during pruning naming_core: - v0.2: few-shot blocklist, stricter is_valid_name (min 3 chars, no digits, no brackets), prompt fragment rejection expanded Miri clarification: the 261 "empty-string" files contain only roads (37) and railroads (37) — infrastructure features never in naming scope. All cities/rivers/oceans/mountains/POIs are clean. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
9ad9b88d7c |
feat(tooling): Gemma 4 batch naming pipeline with wiki-grounded register selection (#833)
Replace the one-at-a-time Gemma 2 naming pipeline with a batch-oriented Gemma 4 E2B pipeline. Key changes: - naming_core.py: shared library with Levenshtein distinctiveness ranking, batch prompt building, mood injection pool, name validation, and adjacent-register refill logic - Wiki-grounded register selection: per-system LLM call picks the cultural register based on wiki/GTTR content instead of hash randomizer - Batch naming: requests N*2 names per call, ranks by word-average Levenshtein distance, fills quota from most-distinct candidates - Mood pool: 13 emotional seeds randomized per-body for vocabulary divergence (ambition, fear, isolation, defiance, etc.) - Adjacent-register refill: when primary register exhausts, automatically switches to next corridor substyle - Inhabited-first body ordering: habitable worlds get first pick of register vocabulary, barren moons get leftovers - Process group cleanup: SIGTERM/SIGKILL the full distrobox chain on subprocess refresh to prevent GPU zombie processes - qa_naming.py: QA report, fix_fewshot_bleed.py: post-hoc fix script - test_batch_naming.py, test_register_selection.py: test harnesses Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
474b31b214 |
fix(tooling): only log rejections when all attempts exhausted (#833)
Individual dedup/blocklist/placeholder/empty rejections that recover on the next attempt are now silent. Only the skipped: summary line prints when all 5 attempts fail. Subprocess errors still print immediately (those indicate a real problem). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
b63c2be8dd |
feat(tooling): 5 retry attempts + refresh every 1000 requests (#833)
Bumped max_attempts from 3 to 5 — with per-system dedup and no stem cap, the remaining dedup hits are mostly per-body collisions which a couple extra attempts with rotated pools can escape. Bumped --refresh default from 200 to 1000. Fewer subprocess restarts = fewer model reloads via distrobox. KV-cache bleed risk is lower now that the validation gauntlet is lighter. Reverted the batch-prompt experiment — Gemma 2 2B drifts on multi-line output; individual calls are more reliable. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
05a1366544 |
refactor(tooling): drop stem cap + scope dedup to per-system (#833)
The stem cap (--stem-cap 20) was rejecting valid names because common feature-type vocabulary tokens like "ridge", "hill", "range" hit the cap after ~200 bodies and blocked all subsequent names containing them. With sub-style rotation already providing variety, the cap was doing more harm than good. Removed entirely. Cross-body dedup narrowed from (hop, corridor, feature_type) to (system_id, feature_type). Two rivers in the same system can't share a name; two rivers in different systems can. This matches how settlers actually name things — they don't coordinate with other star systems. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
cb178a8c57 |
feat(tooling): per-system sub-style rotation across all corridors (#833)
Replace single-inflection corridor palettes with lists of sub-styles. Each system picks one deterministically via hash(system_id), so all bodies in the same system share a cultural register but neighbouring systems get different registers. Core corridor splits into 6 sub-styles (English rural, British colonial, US rural, US cosmopolitan, classical/institutional, Australian/NZ). North/south/east/west reach each get 5 sub-styles covering their cultural spectrum. Deep frontier gets 3 (founder-name, surveyor-descriptive, outpost-functional). This multiplies Gemma's effective vocabulary per corridor by the sub-style count, dramatically reducing dedup pressure. A 6-style core corridor means each sub-style serves ~4 systems instead of 24, so "The Ridge" exhausts after ~4 systems, not ~24. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
3ab28138a0 |
feat(tooling): skip dedup-exhausted features instead of palette fallback (#833)
When all 3 LLM attempts are rejected (dedup, blocklist, etc.), name_feature now returns None instead of a deterministic palette fallback. process_body leaves the name as null in markers.json. The preserved path (_is_blank) treats null as unnamed, so a fill round (re-running the script) picks up only the skipped features with a fresh corpus — zero dedup pressure from the first pass. The fill round can use a different seed, slower prompt, or a different backend entirely (e.g. Haiku). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzerandClaude Opus 4.6
|
c767f82fbb |
feat(tooling): scope cross-body dedup to hop+corridor (#833)
Changed dedup key from (corridor, feature_type) to
(hop, corridor, feature_type). Systems at the same gate-hop distance
in the same corridor are near neighbors and shouldn't share feature
names; systems at different hops can. This prevents corpus exhaustion
where Gemma's narrow range-name distribution ("The Ridge", "Blackwood
Range") collides after ~20 bodies and drives fallback rates toward
100%.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
||
|
jpmschweitzerandClaude Opus 4.6
|
e0707d8072 |
feat(tooling): always log retry reasons in gemma_naming.py (#833)
Retry rejection lines (dedup, blocklist, placeholder, stem_cap, empty,
error) now print unconditionally, not only under --verbose. The
fallback line also includes a tally of the rejection reasons that
exhausted all attempts, e.g.:
fallback: GJ144e-1/range_43 → 'Kirkwood Spine' [blocklist=2 dedup=1]
Diagnostic run on 20 bodies confirms dedup is the primary fallback
driver. Gemma converges on a narrow set of range names ("The Ridge",
"Blackwood Range", "The Spine") that collide across bodies in the
same corridor. Blocklist catches "Thames" and "The Great Divide"
correctly. Zero stem-cap or subprocess-error fallbacks observed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
||
|
jpmschweitzerandClaude Opus 4.6
|
c089557d65 |
feat(tooling): out-of-worktree sr-voice discovery + distrobox wrap + prompt capture (#833)
Three additions unlocked by the gfx1201 ROCm debug session. 1. _find_sr_voice() resolves the default binary path to ~/Projects/settled-reach/binaries/sr-voice-rocm (persistent across worktree lifetimes) with a legacy fallback to the main workdir's cargo target dir. Matches #850's plan to ship platform binaries outside the repo. 2. --distrobox <name> wraps the sr-voice subprocess in `distrobox enter <name> --` when the built binary depends on libs that only exist inside a dev container (libhipblas.so.2 on a Bazzite host). Stdio JSONL protocol flows through unchanged. 3. --dump-prompts PATH captures the attempt-0 prompt for every feature as JSONL without calling an LLM. Force --mock and short-circuit name_feature to return a unique deterministic placeholder. Used to feed the same prompt set to alternate backends (Haiku agent, other models) for offline A/B comparison of naming quality independent of the sampling backend. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
jpmschweitzer
|
a5fbce4cc9 |
feat(tooling): per-system gttr context + cosmopolitan corridor palettes (#833)
Two related quality fixes observed mid-run on Sirius + ACB + Ran:
1) Cosmopolitan corridor palettes. The six corridor inflection labels
were single-culture dominant ("administrative English / Gateway-era",
"British / Australian / Irish", "Korean/Japanese/Taiwanese", etc).
Gemma 2 2B interpreted these as "produce ONLY in this register" and
every core body came out anglophone, every east_reach body came out
East Asian. The real Earth diaspora in the setting is cosmopolitan —
a British surveyor on an east_reach moon still names a river after
their aunt in Dorset. The labels now spell out the dominant register
AND explicitly invite cross-cultural variety so Gemma samples from
the full few-shot pool instead of collapsing to one culture.
2) Per-system gttr context (the big one). The gttr.md files under
wiki/star-systems/<slug>/gttr.md already carry a vivid one-sentence
characterisation of every system — "where the rules live", "forty
years old and still in the draft", "the most connected system in
the Reach", "grandparents owned the land". This is a far stronger
cultural signal than the corridor inflection alone.
New column `star_systems.gttr_hook` stores a pre-extracted 45-word
hook per system. `tooling/db/populate_gttr_hook.py` parses each
gttr.md, regex-matches the first `**NAME**` paragraph, normalises
whitespace, truncates softly at a word cap, and stores it. Covers
all 301 systems (full coverage). Idempotent, safe to re-run after
any wiki update. Explicit transaction wrapper.
gemma_naming.py loads the hook cache at startup via
`load_system_gttr_hooks` and threads `system_hook` plus the system
and body proper names through process_body → name_feature →
_build_prompt. The prompt now carries:
System: <proper_name>. Planet: <body_name>.
About the system: <gttr_hook>
Style: British. Answer: Cooper's Creek
Style: Dutch. Answer: Meijer Beek
...
Real-mode smoke on 10 cases across 4 contrasting systems shows the
hook is doing exactly what it should. Sample output on the same
body_id / local_id pairs:
Tau Ceti (cosmopolitan hub) → Oakham River, Riverwood, Bridle Way
Ran (old-family agricultural) → Hart's Well, Blackwood Ridge
ACB (Lattice Commission seat) → Greenhaven, Rudge Brook
Posto Avançado (PT frontier dead-end) → Rio Preto, Serra de Caxias, Cunha's Cove
Posto Avançado went from "likely-English under the old corridor-only
prompt" to actual Portuguese names with a real Brazilian place stem
(Caxias), because the hook explicitly mentions wave_5 Portuguese
founders and frontier dead-end context. The gttr cultural one-liner
is the single strongest lever available for per-system cohesion —
this was the mono-culture issue observed in the first run, now fixed.
Token cost: ~60-90 extra tokens per prompt (hook + ident line).
Inference slowdown: ~5-10% per call. Acceptable for the quality gain.
Also restores 10 markers.json files that were stale from the aborted
run just killed — they were all core bodies at hop 0-1 which benefit
most from the gttr-context upgrade, so re-running them with the new
prompt is worth the ~3 minutes of re-inference.
|
||
|
jpmschweitzer
|
d811a4f792 |
feat(tooling): prune oversized mountain/river counts + richer fallback palette (#833)
Two fixes from observing the first Gemma 2 batch run on Sirius:
1) Prune oversized feature counts. The upstream terrain pipeline emits
every distinct mountain cluster as a separate `mountain_range` and
every flowing path as a separate `river`. At the atlas generator's
512×256 grid this produced bodies with 40-80 named ranges and
10-15 rivers — noise, not information. A single planet with 48
ridges isn't richer, it's unparseable.
`tooling/planet-gen/prune_atlas_features.py` walks every
`markers.json` under `wiki/star-systems/`, ranks each feature type
by a size proxy, and keeps only the top N:
- mountain_ranges: sorted by `area_cells`, top 8 per body
- rivers: sorted by path length, top 6 per body
- oceans / cities / pois: untouched (already small, or
hand-authored by generate_atlas.py)
Sol (GJ-0) is hardcoded-excluded from pruning so the hand-authored
Earth / Mars / moon content stays untouched.
Each pruned body gets its atlas_* rows re-synced via
`sync_markers_to_db` so the DB mirror stays consistent. Bodies
whose wiki folder has no matching row in `bodies` (14 pre-existing
orphans like GJ1156h-1, GJ34Ah-2, …) are pruned in-file but skip
the DB sync to avoid FK violations on atlas_body_grids.
First run results:
bodies scanned: 2394
bodies pruned: 1513
mountain ranges dropped: 11640
rivers dropped: 1382
Safe to re-run — idempotent when a body is already within the caps.
2) Grounded cosmopolitan fallback palette. When Gemma's 3 retries
all fail (dedup, blocklist, stem-cap, placeholder), the code falls
to `_FALLBACK_STEMS[corridor]`. The old table had 10 stems per
corridor, all Latin-institutional (Meridian, Concord, Prefecture,
Cardinal, Lumen, Foro, Tabula, Vox, Axis, Senatus), which produced
the same-y `Axis Spine / Axis Ridge / Axis Heights / Axis Scarp`
clusters the user flagged on Sirius — exactly the old epic-Latin
register the few-shot pools were rewritten to avoid.
Fallbacks now draw from a 30-45 stem grounded cosmopolitan list
per corridor matching the few-shot pool intent:
- core: 45 stems (Ashfield, Bellview, Cedarbrook,
Fairmont, Ironwood, Kirkwood, Linden, Meridian,
Northfield, Riverside, Westbrook, …)
- north_reach: 40 stems (Ashford, Bellfield, Clifford, Drayton,
Elmhurst, Garner, Holmwood, Kelsworth, …)
- west_reach: 35 stems (Altdorf, Bergfjord, Eikhof, Hoogland,
Järvenpää, Kloosterdam, Nieuwpoort, Sørholm,
Svarteberg, Torsfell, Voorhout, Weserhof, Östby, …)
- east_reach: 35 stems (Aomori, Baektu, Chōshi, Fukagawa,
Hanyang, Izumi, Takamine, Yurigawa, …)
- south_reach: 36 stems (Alves, Brandão, Évora, Gomes, Ribeiro,
Serra, Várzea, Hlanganani, Kilimi, …)
- deep_frontier: 30 stems (Okafor, Stenner, Weller, Kellogg,
Stonebrook, Dustgate, Blackwater, …)
Per-feature suffix lists also expanded (e.g. river suffixes now
include Brook, Stream, Flow, Creek on top of the original Run /
Water / Beck / Rill / Course). Net effect: 300-450 unique fallback
combinations per (corridor, feature_type), up from 50, in the same
grounded register the few-shot pools teach.
Also preserves aliases `inner_corridor`, `inner_orbit`, and
`sol-gateway-axis` as legacy-compatible keys pointing at the
administrative-English palette.
Combined effect on the next run:
- ~45% fewer features to name (pruned 13k/52k)
- ~9× more fallback variety per corridor when fallback does trigger
- Same grounding overhaul from the previous commit, now reaching
into the safety-net path
|
||
|
jpmschweitzer
|
cbeadc3184 |
feat(tooling): grounding overhaul + richer naming palettes (#833)
Substantial quality pass on gemma_naming.py driven by user review of
the first real-mode smoke test output. The earlier run produced names
that read too sci-fi / epic-fantasy / same-y: Aureus, Aetheria,
Stellaris, Nexus, Elysium. Root cause analysis + fixes:
1. Runtime timestamps. The log prefix is now
`[HH:MM:SS +00h03m]` — clock time plus elapsed-since-start. Gives
the user an at-a-glance sense of how long the run has been going
without scrolling back to the banner.
2. System / body headers. When the loop enters a new system it prints
`── SYSTEM K/N GJ 71 — Tau Ceti (hop 0)`. Each body line now
shows `GJ71c (Threshold)` if the body has a proper_name in
systems.db, so the log reads like a tour of the reach rather than
a wall of body_id slugs. Preserved (already-named) bodies now log
a compact "(skip — N names already set)" line so progress is
visible even when no inference happened.
3. Prompt grounding overhaul. The old few-shot examples were all
classical/epic (Wolcott Beck, Nakamura Stream, Ribeiro do Sal,
Drayton Spine) which biased Gemma 2 2B toward Latin/Greek
coinages. New preambles use the shape:
"Settlers named X after themselves, after what they saw, or
after places back home. Most names are mundane, short, and
direct — a surname, a compass direction, a feature, a
practical description. Classical or epic names are rare."
Combined with grounded example pools, Gemma now produces names
like "Cooper's Creek", "Western Ridge", "The Highroad",
"Blackwood Creek", "Dustbowl".
4. Core corridor relabel. The "core" palette inflection was
"institutional Latin / pan-Anglo / Gateway-era", which pattern-
matched in Gemma's training data to "make up Latin-sounding
words" (→ Ardenia, Aurelia, Stellaris). Now it's
"administrative English / Gateway-era" and the outputs are
prosaic — Port Dundas, East Ridge, Meridian, Landing.
5. Rotating few-shot example pools. Each feature type now has 5-7
pools of 5-6 examples each. `_build_prompt()` picks a pool
deterministically per (body_id, local_id, attempt) so:
- Same feature always gets the same prompt (determinism preserved).
- Neighbouring features on the same body get different prompts
(output variance — the sampler doesn't collapse to a single
mode when you ask for 16 mountain names in a row).
- Retries rotate to a new pool, not just a bumped seed, giving
dedup failures a clean second attempt.
6. Cosmopolitan cultural variety in the examples. Earlier pools only
showed British/Australian, Korean/Japanese, Portuguese/Swahili,
German/Dutch/Nordic axes — the four reach corridors. Gemma learned
"names come in four flavours". New pools span Dutch, Nordic,
Italian, French, Polish, Hungarian, Czech, Spanish, Russian,
Finnish, Greek, Irish, Japanese, and British — teaching the model
that names can be any real Earth cultural register, not just the
corridor label. The result: actual Dutch names (Egelantier,
Hochland, Van Damhoeve), actual Nordic (Lundstad, Brygga),
actual Italian (Borgo Marconi, Piazza Nuova), etc.
7. First-name possessive pools. Per user feedback, settler naming
includes both surnames ("Cooper's Creek") and first names
("Clifford's Bay", "Maura's Run", "Yuki's Pool"). Each feature
type now has a dedicated first-name-possessive pool in addition
to the existing surname pool — the two rotate alongside so both
patterns show up without either dominating.
8. One "classical/Latinate" pool per feature type (≈17% of calls
given 5-7 pools per type). Keeps occasional Latin flavour without
making it dominant — the user explicitly noted that replacing
one pattern with another "is never a clean fix for a randomizer."
9. Earth-name blocklist expansion. The Gemma 2 model reached for
real European names ("Weser", "Rhine", "Reykjavik") in the first
real run. Added 21 European rivers (Rhine, Weser, Elbe, Oder,
Vistula, Loire, Rhône, Douro, Tagus, Ebro, Po, Arno, Tiber, …)
and 25 Nordic/Eastern European cities (Reykjavik, Oslo, Gdansk,
Krakow, Prague, Warsaw, Budapest, Belgrade, …). Case-insensitive
"The <name>" stripping still applies so "The Great Divide" also
matches "Great Divide".
Combined smoke test after these changes (10 real-mode prompts across
core + west_reach):
- core: Port Dundas, The Backbone, Dustbowl, Blackwood Creek
- west_reach: Egelantier, Hochland, Der Rücken, Lundstad, Klipfjord
- no placeholder residue, no markdown, no 5+ word outputs.
--shard is gone (dead code since GPU contention killed parallelism).
Resume semantics are still free: re-run the same command and
already-named bodies skip via the preserved path.
|
||
|
jpmschweitzer
|
6decf56b2d |
refactor(tooling): drop --shard from gemma_naming.py (#833)
Parallelism via two concurrent sr-voice subprocesses does not work on this ROCm + llama-cpp-rs setup — launching a second instance poisons the first one's GPU context (both fall back to 0% GPU / 50% CPU busy-loop and stop making progress). Verified empirically: single shard runs cleanly at ~1.2s/feature, two shards deadlock. Without a working parallel path, --shard is dead weight. Resume semantics were already free: the pipeline skips bodies whose markers.json has non-empty name fields (preserved path), so a killed run re-starts just by re-running the same command. Simplifications: - Remove --shard argument and all slicing logic. - Remove banner_shard / shard_offset / shard_n / shard_m plumbing. - Rename internal total_shard_systems → total_systems. - Default --log path is now .tmp/gemma_naming.log (was conditional on --shard). Pass `--log -` to disable file logging. - Startup banner now prints a one-line resume reminder so the user can see at a glance that a killed run is recoverable. |
||
|
jpmschweitzer
|
1efbbcaea9 |
feat(tooling): gemma_naming.py batch naming pipeline for atlas (#833)
New end-to-end pipeline that walks every markers.json in the reach and
fills empty `name` fields using the Gemma 2 voice pipeline via
`sr-voice serve --stdio`. Per D-191 §4: the same Gemma 2 pipeline the
client uses for NPC voicing also produces the atlas content, which is
dual-purposed as a quality test of the LLM plumbing.
Pipeline per body (hop-ordered, core-first):
1. Load markers.json; identify feature records whose `name` is
blank (null or ""). Hand-authored names are never overwritten;
the 6 template bodies and any partial authoring stay put.
2. Look up body context (planet_class, settlement_pattern,
cultural_corridor, population, economic_role) from systems.db.
3. Build a short corridor-aware few-shot prompt per feature type.
Prompts carry 3 concrete `Style: X. Answer: Y` examples so
Gemma 2 2B completes a pattern instead of generating to an
open-ended instruction — this is the single biggest lever
against placeholder echoes on a small model.
4. Stream the prompt into a long-lived sr-voice subprocess, read
the JSONL response, post-process (strip markdown, label
prefixes, brackets, reject 5+ word outputs and placeholder
tokens), check the earth-name blocklist, check per-(corridor,
feature_type) + per-body dedup, check the per-stem cap, retry
up to 3 times with a bumped seed.
5. On persistent failure, fall back to a deterministic palette
generator so every feature ends up with a name.
6. Write markers.json atomically and refresh atlas_* DB rows via
sync_markers_to_db. Commit the DB per body so a crash loses
at most one body of state.
7. Restart the sr-voice subprocess every `--refresh` requests
(default: 200) to prevent KV-cache context bleed.
Core design decisions:
- Determinism: per-(world_seed, body_id, feature_local_id, attempt)
seed so the full run is reproducible.
- Ordering: bodies are processed in ascending `hop_distance_from_gateway`
so core bodies get first pick at every unique Gemma output and
outer sectors fall into the palette fallback when they lose the
dedup race.
- Dedup scope: (cultural_corridor, feature_type) across the run,
PLUS a per-body cross-type set so the same name can't be a river
AND an ocean AND a mountain on the same world. Hand-authored names
are seeded into both sets on load so templates win priority.
- Stem cap: each non-generic root token (e.g. 'Arcturus', 'Meridian')
may appear at most `--stem-cap` times across the full run (default
20), preventing single-word runaway. Fallback names bypass the cap.
- Earth blocklist: 181 curated entries covering major Earth cities,
mountains, rivers, oceans, historical/colonial spellings, and
Greek/Roman mythology that reads too literally. Prefixed variants
('Nouveau Paris', 'New Tokyo') explicitly allowed per the product
intent that Earth-echo names are fine but must not dominate.
Leading 'The ' is stripped before comparison so 'The Great Divide'
also matches.
Operational features:
- `--shard N/M` slices the body list into M partitions for parallel
runs. Two terminals × `--shard 0/2` + `--shard 1/2` fits the
~2.5 GB/instance VRAM footprint twice under the 50% cap on a
16 GB AMD GPU and roughly halves wall time.
- `--log PATH` writes a timestamped tee of every status line to a
file. Default: `.tmp/gemma_naming.shard{N}of{M}.log` when a
non-trivial shard is in use.
- SQLite `PRAGMA journal_mode=WAL` + `busy_timeout=15000` so two
concurrent shards serialize writes without lock errors.
- Per-body progress lines report `body K/N`, `sys K/N`, and
`hop=H` so the user can watch core sectors finish first.
- Each body logs the new names it produced per feature type so the
user can eyeball quality as the run progresses.
- Checkpoint summary every 25 bodies: cumulative names, rate,
ETA — gives the log regular scroll points.
- `--mock` uses `server/sr-voice/mock-stdio.sh` for dry-fire
pipeline validation without a model load (tested end-to-end).
Supporting files:
- `tooling/planet-gen/earth_blocklist.txt` — 181 curated entries.
- `tooling/db/backfill_cultural_corridor.py` — one-off migration
that fills the `cultural_corridor` column on both `star_systems`
and `bodies` from the `geographic_sector` values. Before this
pass, 99.4% of rows (3221/3240) had a NULL cultural_corridor
despite `wiki_sync.py` being aware of the column — the wiki
index.md files only carry the sector header, which was never
propagated to the DB column. Idempotent, safe to re-run after
any wiki_sync rebuild, explicit transaction wrapper with
rollback on failure.
Full batch runtime estimate: ~20 hours single-shard / ~10 hours
double-shard on this hardware. Smoke tests across five hardened
iterations (v1–v5) on GJ71b/c/d/d-1/e confirm the pipeline produces
clean, varied, culturally-coherent names with zero post-processing
residue.
|