The oracle-validated route DB (the realistic path to 100% under prose-only)
lived only in the gitignored real.db, so a navdata rebuild would lose it.
Now it is portable and version-controlled:
- routes::export_validated -> TSV of every ifps_ok route (one per pair);
import_seeds loads them back as reusable seeds (source=seed).
- CLI `export-routes` / `import-routes`.
- discover reuse now matches any ifps_ok route (not just source=ifps), so
imported seeds are reused instantly (verified round-trip).
- data/validated_routes.tsv: 65 validated European city-pair routes, built
by bulk-seeding discover over ~85 common pairs (the DB grew ~11 -> 65).
This is the PFPX-style route DB, committed: 65 pairs are now no-error &
instant on any machine; `learn` + more bulk-seed grow it further.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Completes the route-DB pipeline that is the realistic path to 100% under
the prose-only / independent constraint (pure RAD-prose synthesis can't
guarantee 100% on arbitrary pairs):
- discover: synthesize -> oracle-validate -> store if no-error -> reuse
instantly next time (verified: 2nd LFPG-EDDF returns via "reused stored
IFPS-valid route", 0 IFPUV calls).
- learn --from --to --route --fl: validate a user-supplied route against
IFPUV and, if no-error, record it (source=ifps, ifps_ok=1) so discover
reuses it forever. --force keeps a still-erroring route for review
(source=manual, not reused). This is how a real ops route DB grows and
how the pairs the generator can't yet crack reach 100% (seed once).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The transition parser dropped airways, so mandate splices filed all-DCT
(MOVUM DCT HAREM) — wrong routing, and the arrival splice self-looped
(ROUTE49) when the enroute already crossed the arrival area.
- rad: expand_via keeps airway tokens (is_airway_tok); a transition is now
a point/airway token sequence (DCT where no airway), e.g.
[MOVUM, T109, HAREM, T104, WLD].
- discover: transition_legs() builds legs with the interspersed airways.
route_via_arrival_transition anchors on the transition ENTRY (first
on-airway point) and flies the transition through once — the enroute leg
stops before the arrival area, so no self-loop. route_via_transition and
the est/near sorts use the last/first *point* (skip airways).
Files the mandate correctly now (… MOVUM T109 HAREM T104 WLD DCT EDDM).
Coverage 6/10 held, no regression (LFPG-EDDF still passes); LSZH-EGLL
6→4 PROF204; EGLL-EDDM's fake 1-err self-loop is replaced by its real
residuals (CPT dep forbidden, forbidden points on the mandated T104, WLD
STAR limit). 49 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Parse Annex 2B mandatory arrival routings and repair the best few
candidates (not just the lowest-error one):
- core::rad: MandatoryArr + RadData.mandatory_arr +
mandatory_arr_transitions(ades) -> allowed point-sequences.
- rad crate: parse_mandatory_arr() (VIA-clauses in NOT-AVBL-EXC rows,
scoped to ARR airports). via_sequences() now scans all parenthesised
transition groups (contain DCT/airway), not just contiguous "VIA (",
because sub-option markers separate them ("VIA i) (KARDU DCT KALMO)").
rad-tool `manarr [ADES]`. LIML now yields KARDU KALMO etc.
- discover: "manarr" candidates route to the transition's on-airway
anchor then splice the (often off-airway) transition points in via DCT.
- attempt_once repairs the top-2 candidates and keeps the best result:
the lowest-raw-error candidate isn't always the most repairable (a
mandate-satisfying route may carry more mechanical errors than a route
stuck on an unfixable mandatory).
Collapses the dense-FRA near-misses: EGLL-EDDM 7->1 err, EDDF-LIRF 3->1
(mandatory PROF205 resolved; residual is a single ROUTE49). Coverage 6/10,
no regression (passing pairs short-circuit on accept). 49 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Parse Annex 3A DEP "COMPULSORY FOR TFC" rows into structured mandatory
departure routings and satisfy PROF205 "… ANNEX3A DEP … MANDATORY ROUTE"
by offering the allowed transitions as discovery candidates:
- core::rad: MandatoryDep + RadData.mandatory_dep +
mandatory_dep_transitions(adep) -> allowed point-sequences.
- rad crate: parse_mandatory_dep() extracts the VIA(...) clauses, expands
nested "(A, B)" alternatives into separate branches, and chains
sub-fragments into full sequences from an entry point (46 rules; LFPG =
20 transitions incl. NURMO CMB VEKIN ADUTO, OPALE KESAX DIMAL ALESO).
rad-tool `mandep [ADEP]` to inspect.
- discover: for the departure airport, inject "mandep" candidates that fly
the allowed transitions nearest the destination (additive — the oracle
still picks best, so no regression risk).
Cracks LFPG-EDDF (Paris-Frankfurt): coverage 5/10 -> 6/10 on the batch,
no regression. 49 tests green. Phase 2b (mandatory arrivals) is next.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Cracks EDDM-LOWW (coverage 4/10 -> 5/10) and generalises DCT-too-long
(ROUTE165) repair:
- airway_path_avoiding(): the splice now builds the avoid-aware graph, so
re-routing a too-long DCT never re-introduces a RAD-forbidden airway.
- Off-airway absorb: when a broken segment's from-point isn't on any airway
(a SID/terminal fix, e.g. ALUTU), splice from the previous on-airway
point through to the target, dropping the DCT-only detour
(EDDM …BIBAG DCT ALUTU DCT AGNAV -> BIBAG airway AGNAV).
- PROF204 repair now bans the *segment* X<airway>Y (seg_key) instead of
excluding the points, so a forbidden airway leg out of a SID/STAR gateway
(BIBAG) no longer nukes the gateway itself.
Batch: 5/10, no regression. 49 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
If discovery's upper-airspace route can't be made no-error because of RAD
restrictions we don't yet model (PROF204 forbidden / PROF205 mandatory
routing common in Free Route Airspace), retry the whole discovery in the
conventional airway regime below the FRA floor (~FL195), where those
restrictions are far sparser. Adopt the low route only if it is fully
valid — the bar is no-error, and flying needlessly low isn't the answer
when neither passes, so the natural-FL route is kept otherwise.
Measured on the batch: no regression, and the fallback collapses the
FRA-blocked pairs from ~7-9 errors to 1-2 (near-miss). It delivers a
no-error plan for any pair that has a valid low-level alternative; the
remaining near-misses need the phase-2 mandatory-routing engine to close.
49 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Parse the 1135 forbidden airway segments in RAD Annex 2B (rows with
Airway/From/To + "NOT AVBL FOR TFC") into structured ForbiddenSeg records
(FL band + arrival/departure applicability), and route around the ones
that apply to a flight from the start:
- core::rad: ForbiddenSeg + RadData.forbidden_segs + forbidden_segments(
adep, ades, fl) -> applicable (airway,from,to) triples.
- rad crate: parse_forbidden_segments() wired into parse(); rad-tool
`forbidden [ADEP ADES FL]` to inspect.
- graph: seg_key() encodes a segment ban into the existing avoid channel;
build_avoiding skips those segments. discover seeds the avoid set with
the flight's Annex-2B bans before generating the airway candidate.
Measured: no regression (coverage 4/10 unchanged on the batch) — this is
the *forbidden* half. The coverage mover is mandatory routing (PROF205:
"must route via X"), which pure avoidance can't provide; that's phase 2.
49 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two correctness fixes exposed by a coverage batch where "1 ROUTE130"
near-misses turned out to be catastrophically broken routes (Paris->
Frankfurt routed via Alaska) hidden by IFPS error-masking:
- airway_path now rejects a spliced sub-path that balloons past
SPLICE_MAX_RATIO x the direct distance. At a fragmented FL the graph
could route two nearby fixes the long way around the whole network; that
garbage then showed only 1 (masked) error and beat sane routes in the
monotonic best-tracking.
- route_item15 strips computer-navigation fixes (unnamed ARINC waypoints
whose ident carries a digit, e.g. GT27A) proactively, collapsing
same-airway legs, so IFPS never sees a ROUTE130 designator it would
reject and stop evaluating at.
Net: coverage numbers are now honest (masked garbage no longer counts as a
near-miss); the real blocker is RAD PROF204/205 on upper/FRA routes.
49 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Graph gains build_avoiding(): drop segments whose endpoint ident or
airway is in an avoid set. plan_route_conn_avoiding /
plan_route_best_avoiding thread it through (also from SID/STAR conns).
- New PROF204 repair: parse "TRAFFIC VIA <pts> IS ON FORBIDDEN ROUTE",
exclude the cited enroute points (never airports), re-plan around them.
Runs first (a hard "cannot use" beats splice/level moves) and the avoid
set accumulates across iterations. Other re-plans are now avoid-aware.
- Repair loop tolerates MAX_STALL non-improving steps before reverting, so
it can cross a valley where excluding one forbidden point forces a
temporarily worse route en route to a fix. best-so-far still guards the
result: this can only find an equal-or-better route, never a worse one.
LSZH-LFBZ: 13 -> 2 errors (FRA-border problem gone; remainder is one
Annex-2B-forbidden Swiss departure segment needing real SID modelling).
LFRS-LFST / LFRN-LFMN / LSGG-LFPO still pass, 0 iterations. 49 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three router bugs surfaced by testing real filed pairs against the
Eurocontrol IFPUV oracle:
- ROUTE130 repair now collapses the leg chain when dropping an unknown
designator (a computer-nav fix like GT27A on an airway), instead of
leaving a dangling `AWY AWY`. Same airway on both sides -> merge; else DCT.
- FL detour ratio 1.6 -> 1.25: the FL-filtered graph fragments the
conventional (<=FL195) network and forced an absurd London detour for a
domestic French hop; the short full route now wins.
- FL ceiling-cap: file a short low-level pair at its airways' natural
ceiling instead of an optimistic FL360 (which yields PROF195 on every
leg). Gated on upper-network fragmentation so genuine high-level
long-haul (EGLL-LSZH) is never wrongly lowered.
- New PROF195 repair lowers the cruise FL to where the cited airway exists.
LFRS-LFST now no-error at FL080 (0 iterations); LFRN-LFMN and LSGG-LFPO
still pass and now converge in 0 iterations. 4 new unit tests; 49 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Clean-room reproduction of PFPX: route generation + IFPS validation + OFP.
Independent design — official ICAO/EUROCONTROL data only (RAD, FRA points,
IFPUV oracle); no community FPL sources.
Workspace crates: core (routing/discover/rad/navdata/perf/export),
cli, server, rad (Annex parser), gui (Tauri v2 + React + MapLibre).
Route discovery: oracle-in-the-loop repair against Eurocontrol IFPUV.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Alexandre